{
  "schemaVersion": 1,
  "engineVersion": "1.2.9",
  "devices": [
    {
      "id": "apple-silicon",
      "title": "Apple Silicon Mac",
      "memory": 64,
      "unit": "Mac memory",
      "adjustable": true,
      "quick": 2,
      "subtitle": "Choose your memory"
    },
    {
      "id": "macbook-neo",
      "title": "MacBook Neo",
      "target": "apple-silicon",
      "memory": 8,
      "unit": "Unified memory",
      "quick": 1,
      "subtitle": "8 GB",
      "note": "MacBook Neo has 8 GB. The smallest Mac recipe is a possible fit. A launch on this laptop is still unverified.",
      "hardwareSource": "https://www.apple.com/macbook-neo/specs/"
    },
    {
      "id": "dgx-spark",
      "title": "One DGX Spark",
      "memory": 128,
      "unit": "Unified memory",
      "quick": 3,
      "subtitle": "128 GB model"
    },
    {
      "id": "dgx-spark-pair",
      "title": "Two DGX Sparks",
      "memory": 128,
      "unit": "Unified memory per Spark",
      "cluster": true,
      "quick": 4,
      "subtitle": "128 GB each",
      "note": "These guides start the model on both Sparks. Connect the two devices before running the setup."
    },
    {
      "id": "intel-arc-b70",
      "title": "Intel Arc Pro B70",
      "memory": 32,
      "unit": "GPU memory",
      "quick": 5,
      "subtitle": "32 GB · Linux",
      "note": "These Intel recipes use the Linux build. Follow the build guide before copying a command.",
      "hardwareSource": "https://www.intel.com/content/www/us/en/products/sku/245797/intel-arc-pro-b70-graphics/specifications.html"
    },
    {
      "id": "intel-iris-xe",
      "title": "Intel Iris Xe · Windows with WSL2",
      "memory": 8,
      "unit": "Memory assigned to WSL2",
      "note": "Tested on a Gen12LP Iris Xe laptop with Ubuntu 24.04 running in WSL2 and 8 GB assigned to WSL."
    },
    {
      "id": "dgx-spark-64",
      "title": "One DGX Spark · 64 GB",
      "target": "dgx-spark",
      "memory": 64,
      "unit": "Unified memory",
      "knownMemoryOnly": true,
      "note": "These are memory estimates for the 64 GB model. The published Spark launch test used a 128 GB device.",
      "hardwareSource": "https://www.nvidia.com/en-us/products/workstations/dgx-spark/"
    },
    {
      "id": "cuda",
      "title": "Other NVIDIA GPU",
      "memory": 24,
      "unit": "GPU memory (VRAM)",
      "adjustable": true
    },
    {
      "id": "dgx-spark-trio",
      "title": "Three DGX Sparks",
      "memory": 128,
      "unit": "Unified memory per Spark",
      "cluster": true,
      "menu": "DGX Spark clusters"
    },
    {
      "id": "dgx-spark-quad",
      "title": "Four DGX Sparks",
      "memory": 128,
      "unit": "Unified memory per Spark",
      "cluster": true,
      "menu": "DGX Spark clusters"
    },
    {
      "id": "rtx-pro-6000",
      "title": "One RTX PRO 6000",
      "memory": 96,
      "unit": "GPU VRAM",
      "menu": "RTX PRO 6000 Blackwell"
    },
    {
      "id": "rtx-pro-6000-pair",
      "title": "Two RTX PRO 6000s",
      "memory": 96,
      "unit": "VRAM per GPU",
      "menu": "RTX PRO 6000 Blackwell"
    },
    {
      "id": "rtx-pro-6000-quad",
      "title": "Four RTX PRO 6000s",
      "memory": 96,
      "unit": "VRAM per GPU",
      "menu": "RTX PRO 6000 Blackwell"
    },
    {
      "id": "amd-hip",
      "title": "Strix Halo · Radeon 8060S",
      "memory": 128,
      "unit": "Unified memory"
    }
  ],
  "backends": [
    {
      "id": "metal",
      "title": "Metal",
      "devices": [
        "apple-silicon",
        "macbook-neo"
      ]
    },
    {
      "id": "cuda",
      "title": "CUDA",
      "devices": [
        "dgx-spark",
        "dgx-spark-pair",
        "dgx-spark-64",
        "cuda",
        "dgx-spark-trio",
        "dgx-spark-quad",
        "rtx-pro-6000",
        "rtx-pro-6000-pair",
        "rtx-pro-6000-quad"
      ]
    },
    {
      "id": "hip",
      "title": "AMD HIP",
      "devices": [
        "amd-hip"
      ]
    },
    {
      "id": "xpu",
      "title": "Intel XPU",
      "devices": [
        "intel-arc-b70",
        "intel-iris-xe"
      ]
    }
  ],
  "records": [
    {
      "id": "qwen3.5-2b",
      "name": "Qwen3.5-2B",
      "provider": "Qwen",
      "mark": "qwen-color.svg",
      "description": "A lightweight text model for getting started.",
      "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/qwen3.5-2b-native.md",
      "test": "curl -fsS http://127.0.0.1:8080/v1/chat/completions \\\n  -H 'Content-Type: application/json' \\\n  -d '{\"model\":\"qwen3.5-2b\",\"messages\":[{\"role\":\"user\",\"content\":\"Say hello in one sentence.\"}],\"max_tokens\":64}'",
      "deployments": {
        "apple-silicon": {
          "format": "MLX 4-bit",
          "estimateGB": 8,
          "weightGB": 1.7,
          "checkpoint": "https://huggingface.co/mlx-community/Qwen3.5-2B-MLX-4bit/tree/93760be4f1f69842a46bc13dbdc0f19e291392a3",
          "runtime": "Native Metal",
          "setup": "Install the TensorFold CLI on your Mac to download the pinned weights and start the server.",
          "setupUrl": "https://github.com/ashhart/TensorFold/blob/v1.2.9/README.md#mac-with-apple-silicon",
          "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/qwen3.5-2b-native.md",
          "engineVersion": "1.2.9",
          "context": 8192,
          "parallel": 1,
          "settings": {
            "defaults": {
              "context": 8192,
              "parallel": 1,
              "input": "text"
            },
            "contexts": [
              8192,
              16384,
              32768,
              65536
            ],
            "requests": [
              1,
              2,
              4,
              8
            ],
            "inputs": [
              "text"
            ]
          },
          "command": "(\n  set -e\n  export HF_HUB_CACHE=\"${HF_HUB_CACHE:-${HF_HOME:-$HOME/.cache/huggingface}/hub}\"\n  tensorfold pull mlx-community/Qwen3.5-2B-MLX-4bit@93760be4f1f69842a46bc13dbdc0f19e291392a3\n\n  tensorfold serve \\\n    \"$HF_HUB_CACHE/models--mlx-community--Qwen3.5-2B-MLX-4bit/snapshots/93760be4f1f69842a46bc13dbdc0f19e291392a3\" \\\n    --name qwen3.5-2b --host 127.0.0.1 --port 8080 \\\n    --context 8192 --parallel 1\n)"
        },
        "dgx-spark": {
          "format": "MLX 4-bit",
          "estimateGB": 8,
          "weightGB": 1.7,
          "checkpoint": "https://huggingface.co/mlx-community/Qwen3.5-2B-MLX-4bit/tree/93760be4f1f69842a46bc13dbdc0f19e291392a3",
          "runtime": "Podman · CUDA",
          "setup": "Requires Podman, NVIDIA driver R580+ and NVIDIA CDI. The first run builds the image and caches the weights.",
          "setupUrl": "https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/cdi-support.html",
          "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/qwen3.5-2b-native.md",
          "engineVersion": "1.2.9",
          "context": 8192,
          "parallel": 1,
          "settings": {
            "defaults": {
              "context": 8192,
              "parallel": 1,
              "input": "text"
            },
            "contexts": [
              8192,
              16384,
              32768
            ],
            "requests": [
              1,
              2,
              4
            ],
            "inputs": [
              "text"
            ]
          },
          "command": "(\n  set -e\n  image=localhost/tensorfold:1.2.9\n  podman image exists \"$image\" || \\\n    podman build -t \"$image\" --build-arg TENSORFOLD_VERSION=1.2.9 \\\n      'https://github.com/ashhart/TensorFold.git#6b9de695104b6b8b30c20458da87ac06c9480289'\n\n  podman run --rm --pull=never -v tensorfold-weights:/huggingface:U \\\n    \"$image\" pull mlx-community/Qwen3.5-2B-MLX-4bit@93760be4f1f69842a46bc13dbdc0f19e291392a3\n\n  podman run --rm --pull=never --device nvidia.com/gpu=0 \\\n    -p 127.0.0.1:8080:8080 -v tensorfold-weights:/huggingface:ro \\\n    \"$image\" serve \\\n    \"/huggingface/hub/models--mlx-community--Qwen3.5-2B-MLX-4bit/snapshots/93760be4f1f69842a46bc13dbdc0f19e291392a3\" \\\n    --name qwen3.5-2b --host 0.0.0.0 --port 8080 \\\n    --context 8192 --parallel 1\n)"
        },
        "cuda": {
          "format": "MLX 4-bit",
          "estimateGB": 8,
          "weightGB": 1.7,
          "checkpoint": "https://huggingface.co/mlx-community/Qwen3.5-2B-MLX-4bit/tree/93760be4f1f69842a46bc13dbdc0f19e291392a3",
          "runtime": "Podman · CUDA",
          "setup": "Requires Podman, NVIDIA driver R580+ and NVIDIA CDI. The first run builds the image and caches the weights.",
          "setupUrl": "https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/cdi-support.html",
          "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/qwen3.5-2b-native.md",
          "engineVersion": "1.2.9",
          "context": 8192,
          "parallel": 1,
          "settings": {
            "defaults": {
              "context": 8192,
              "parallel": 1,
              "input": "text"
            },
            "contexts": [
              8192,
              16384,
              32768
            ],
            "requests": [
              1,
              2,
              4
            ],
            "inputs": [
              "text"
            ]
          },
          "command": "(\n  set -e\n  image=localhost/tensorfold:1.2.9\n  podman image exists \"$image\" || \\\n    podman build -t \"$image\" --build-arg TENSORFOLD_VERSION=1.2.9 \\\n      'https://github.com/ashhart/TensorFold.git#6b9de695104b6b8b30c20458da87ac06c9480289'\n\n  podman run --rm --pull=never -v tensorfold-weights:/huggingface:U \\\n    \"$image\" pull mlx-community/Qwen3.5-2B-MLX-4bit@93760be4f1f69842a46bc13dbdc0f19e291392a3\n\n  podman run --rm --pull=never --device nvidia.com/gpu=0 \\\n    -p 127.0.0.1:8080:8080 -v tensorfold-weights:/huggingface:ro \\\n    \"$image\" serve \\\n    \"/huggingface/hub/models--mlx-community--Qwen3.5-2B-MLX-4bit/snapshots/93760be4f1f69842a46bc13dbdc0f19e291392a3\" \\\n    --name qwen3.5-2b --host 0.0.0.0 --port 8080 \\\n    --context 8192 --parallel 1\n)"
        }
      },
      "capabilities": {
        "use": "Start with classification, short rewrites and simple extraction.",
        "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/qwen3.5-2b-native.md",
        "epoch": null,
        "size": "2B",
        "sizeLabel": "2 billion parameters",
        "sizeSource": "https://huggingface.co/Qwen/Qwen3.5-2B",
        "shortUse": "Short text tasks"
      }
    },
    {
      "id": "lfm2.5-24b-a2b",
      "name": "LFM2.5-24B-A2B",
      "provider": "Liquid AI",
      "mark": "liquid.svg",
      "description": "A mixture-of-experts text model with 2B active parameters.",
      "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/lfm2.5-24b-a2b.md",
      "test": "curl -fsS http://127.0.0.1:8080/v1/chat/completions \\\n  -H 'Content-Type: application/json' \\\n  -d '{\"model\":\"lfm2.5-24b-a2b\",\"messages\":[{\"role\":\"user\",\"content\":\"Say hello in one sentence.\"}],\"max_tokens\":64}'",
      "deployments": {
        "apple-silicon": {
          "format": "MLX 4-bit",
          "estimateGB": 24,
          "weightGB": 13.6,
          "checkpoint": "https://huggingface.co/LiquidAI/LFM2.5-24B-A2B-MLX-4bit/tree/667408d494bde191c2cb9dd4f4549827816b9275",
          "runtime": "Native Metal",
          "setup": "Install the TensorFold CLI on your Mac to download the pinned weights and start the server.",
          "setupUrl": "https://github.com/ashhart/TensorFold/blob/v1.2.9/README.md#mac-with-apple-silicon",
          "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/lfm2.5-24b-a2b.md",
          "engineVersion": "1.2.9",
          "context": 8192,
          "parallel": 1,
          "settings": {
            "defaults": {
              "context": 8192,
              "parallel": 1,
              "input": "text"
            },
            "contexts": [
              8192,
              16384,
              32768,
              40960
            ],
            "requests": [
              1,
              2,
              4,
              8
            ],
            "inputs": [
              "text"
            ]
          },
          "command": "(\n  set -e\n  export HF_HUB_CACHE=\"${HF_HUB_CACHE:-${HF_HOME:-$HOME/.cache/huggingface}/hub}\"\n  tensorfold pull LiquidAI/LFM2.5-24B-A2B-MLX-4bit@667408d494bde191c2cb9dd4f4549827816b9275\n\n  tensorfold serve \\\n    \"$HF_HUB_CACHE/models--LiquidAI--LFM2.5-24B-A2B-MLX-4bit/snapshots/667408d494bde191c2cb9dd4f4549827816b9275\" \\\n    --name lfm2.5-24b-a2b --host 127.0.0.1 --port 8080 \\\n    --context 8192 --parallel 1\n)"
        },
        "dgx-spark": {
          "format": "MLX 4-bit",
          "estimateGB": 24,
          "weightGB": 13.6,
          "checkpoint": "https://huggingface.co/LiquidAI/LFM2.5-24B-A2B-MLX-4bit/tree/667408d494bde191c2cb9dd4f4549827816b9275",
          "runtime": "Podman · CUDA",
          "setup": "Requires Podman, NVIDIA driver R580+ and NVIDIA CDI. The first run builds the image and caches the weights.",
          "setupUrl": "https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/cdi-support.html",
          "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/lfm2.5-24b-a2b.md",
          "engineVersion": "1.2.9",
          "context": 8192,
          "parallel": 1,
          "settings": {
            "defaults": {
              "context": 8192,
              "parallel": 1,
              "input": "text"
            },
            "contexts": [
              8192,
              16384,
              32768
            ],
            "requests": [
              1,
              2,
              4,
              8
            ],
            "inputs": [
              "text"
            ]
          },
          "command": "(\n  set -e\n  image=localhost/tensorfold:1.2.9\n  podman image exists \"$image\" || \\\n    podman build -t \"$image\" --build-arg TENSORFOLD_VERSION=1.2.9 \\\n      'https://github.com/ashhart/TensorFold.git#6b9de695104b6b8b30c20458da87ac06c9480289'\n\n  podman run --rm --pull=never -v tensorfold-weights:/huggingface:U \\\n    \"$image\" pull LiquidAI/LFM2.5-24B-A2B-MLX-4bit@667408d494bde191c2cb9dd4f4549827816b9275\n\n  podman run --rm --pull=never --device nvidia.com/gpu=0 \\\n    -p 127.0.0.1:8080:8080 -v tensorfold-weights:/huggingface:ro \\\n    \"$image\" serve \\\n    \"/huggingface/hub/models--LiquidAI--LFM2.5-24B-A2B-MLX-4bit/snapshots/667408d494bde191c2cb9dd4f4549827816b9275\" \\\n    --name lfm2.5-24b-a2b --host 0.0.0.0 --port 8080 \\\n    --context 8192 --parallel 1\n)",
          "decode": {
            "rate": "87",
            "hardware": "One DGX Spark",
            "conditions": "One greedy decode stream on the 4-bit checkpoint.",
            "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/lfm2.5-24b-a2b.md#cuda",
            "credit": "TensorFold measurements"
          }
        },
        "cuda": {
          "format": "MLX 4-bit",
          "estimateGB": 24,
          "weightGB": 13.6,
          "checkpoint": "https://huggingface.co/LiquidAI/LFM2.5-24B-A2B-MLX-4bit/tree/667408d494bde191c2cb9dd4f4549827816b9275",
          "runtime": "Podman · CUDA",
          "setup": "Requires Podman, NVIDIA driver R580+ and NVIDIA CDI. The first run builds the image and caches the weights.",
          "setupUrl": "https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/cdi-support.html",
          "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/lfm2.5-24b-a2b.md",
          "engineVersion": "1.2.9",
          "context": 8192,
          "parallel": 1,
          "settings": {
            "defaults": {
              "context": 8192,
              "parallel": 1,
              "input": "text"
            },
            "contexts": [
              8192,
              16384,
              32768
            ],
            "requests": [
              1,
              2,
              4,
              8
            ],
            "inputs": [
              "text"
            ]
          },
          "command": "(\n  set -e\n  image=localhost/tensorfold:1.2.9\n  podman image exists \"$image\" || \\\n    podman build -t \"$image\" --build-arg TENSORFOLD_VERSION=1.2.9 \\\n      'https://github.com/ashhart/TensorFold.git#6b9de695104b6b8b30c20458da87ac06c9480289'\n\n  podman run --rm --pull=never -v tensorfold-weights:/huggingface:U \\\n    \"$image\" pull LiquidAI/LFM2.5-24B-A2B-MLX-4bit@667408d494bde191c2cb9dd4f4549827816b9275\n\n  podman run --rm --pull=never --device nvidia.com/gpu=0 \\\n    -p 127.0.0.1:8080:8080 -v tensorfold-weights:/huggingface:ro \\\n    \"$image\" serve \\\n    \"/huggingface/hub/models--LiquidAI--LFM2.5-24B-A2B-MLX-4bit/snapshots/667408d494bde191c2cb9dd4f4549827816b9275\" \\\n    --name lfm2.5-24b-a2b --host 0.0.0.0 --port 8080 \\\n    --context 8192 --parallel 1\n)"
        }
      },
      "capabilities": {
        "use": "Start with local chat, document summaries and repeated short requests.",
        "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/lfm2.5-24b-a2b.md",
        "epoch": null,
        "size": "24B",
        "sizeLabel": "24 billion parameters",
        "sizeSource": "https://huggingface.co/LiquidAI/LFM2.5-24B-A2B",
        "shortUse": "Chat and summaries",
        "sizeDetails": "Uses 2B parameters at a time while generating text."
      }
    },
    {
      "id": "qwen3.5-9b",
      "name": "Qwen3.5-9B",
      "provider": "Qwen",
      "mark": "qwen-color.svg",
      "description": "A 9B text model for a smaller Mac setup.",
      "source": "https://github.com/ashhart/TensorFold/releases/tag/v1.2.2",
      "test": "curl -fsS http://127.0.0.1:8080/v1/chat/completions \\\n  -H 'Content-Type: application/json' \\\n  -d '{\"model\":\"qwen3.5-9b\",\"messages\":[{\"role\":\"user\",\"content\":\"Say hello in one sentence.\"}],\"max_tokens\":64}'",
      "deployments": {
        "apple-silicon": {
          "format": "MLX 4-bit",
          "estimateGB": 16,
          "weightGB": 6,
          "checkpoint": "https://huggingface.co/mlx-community/Qwen3.5-9B-4bit/tree/8b2b98c00a6b4d291155e4890773ca8f769aee53",
          "runtime": "Native Metal",
          "setup": "Install the TensorFold CLI on Apple Silicon. TensorFold’s 1.2.2 release documents native Metal support for the 9B model.",
          "setupUrl": "https://github.com/ashhart/TensorFold/blob/v1.2.9/README.md#mac-with-apple-silicon",
          "source": "https://github.com/ashhart/TensorFold/releases/tag/v1.2.2",
          "engineVersion": "1.2.9",
          "context": 8192,
          "parallel": 1,
          "settings": {
            "defaults": {
              "context": 8192,
              "parallel": 1,
              "input": "text"
            },
            "contexts": [
              8192,
              16384,
              32768
            ],
            "requests": [
              1
            ],
            "inputs": [
              "text"
            ]
          },
          "command": "(\n  set -e\n  export HF_HUB_CACHE=\"${HF_HUB_CACHE:-${HF_HOME:-$HOME/.cache/huggingface}/hub}\"\n  tensorfold pull mlx-community/Qwen3.5-9B-4bit@8b2b98c00a6b4d291155e4890773ca8f769aee53\n\n  tensorfold serve \\\n    \"$HF_HUB_CACHE/models--mlx-community--Qwen3.5-9B-4bit/snapshots/8b2b98c00a6b4d291155e4890773ca8f769aee53\" \\\n    --name qwen3.5-9b --host 127.0.0.1 --port 8080 \\\n    --context 8192 --parallel 1\n)"
        }
      },
      "capabilities": {
        "use": "Start with short summaries, drafting and everyday text questions on a smaller Mac.",
        "source": "https://huggingface.co/Qwen/Qwen3.5-9B",
        "epoch": {
          "name": "Qwen3.5-9B",
          "score": 139.48,
          "low": 137.26,
          "high": 141.69
        },
        "size": "9B",
        "sizeLabel": "9 billion parameters",
        "sizeSource": "https://huggingface.co/Qwen/Qwen3.5-9B",
        "shortUse": "Writing and summaries"
      }
    },
    {
      "id": "nemotron-3.5-lightning",
      "name": "Nemotron 3.5 Lightning",
      "provider": "NVIDIA",
      "mark": "nvidia-color.svg",
      "description": "A 30B mixture-of-experts model with its MTP draft head included.",
      "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/nemotron-3.5.md",
      "test": "curl -fsS http://127.0.0.1:8080/v1/chat/completions \\\n  -H 'Content-Type: application/json' \\\n  -d '{\"model\":\"nemotron-3.5-lightning\",\"messages\":[{\"role\":\"user\",\"content\":\"Say hello in one sentence.\"}],\"max_tokens\":64}'",
      "deployments": {
        "apple-silicon": {
          "format": "MLX 4-bit · MTP",
          "estimateGB": 28,
          "weightGB": 18.5,
          "checkpoint": "https://huggingface.co/TensorFold/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit/tree/d9d758fb83953437f7263256b0d96157e2a348b8",
          "runtime": "Native Metal",
          "setup": "The engine docs cover Apple Silicon, DGX Spark, RTX 3090 and RTX PRO 6000. On CUDA, use NVIDIA driver R580+ and CDI.",
          "setupUrl": "https://github.com/ashhart/TensorFold/blob/v1.2.9/README.md#mac-with-apple-silicon",
          "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/nemotron-3.5.md",
          "engineVersion": "1.2.9",
          "context": 8192,
          "parallel": 1,
          "settings": {
            "defaults": {
              "context": 8192,
              "parallel": 1,
              "input": "text"
            },
            "contexts": [
              8192,
              16384,
              32768
            ],
            "requests": [
              1,
              2,
              4,
              8
            ],
            "inputs": [
              "text"
            ]
          },
          "command": "(\n  set -e\n  export HF_HUB_CACHE=\"${HF_HUB_CACHE:-${HF_HOME:-$HOME/.cache/huggingface}/hub}\"\n  tensorfold pull TensorFold/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit@d9d758fb83953437f7263256b0d96157e2a348b8\n\n  tensorfold serve \\\n    \"$HF_HUB_CACHE/models--TensorFold--NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit/snapshots/d9d758fb83953437f7263256b0d96157e2a348b8\" \\\n    --name nemotron-3.5-lightning --host 127.0.0.1 --port 8080 \\\n    --context 8192 --parallel 1\n)"
        },
        "dgx-spark": {
          "format": "MLX 4-bit · MTP",
          "estimateGB": 28,
          "weightGB": 18.5,
          "checkpoint": "https://huggingface.co/TensorFold/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit/tree/d9d758fb83953437f7263256b0d96157e2a348b8",
          "runtime": "Podman · CUDA",
          "setup": "The engine docs cover Apple Silicon, DGX Spark, RTX 3090 and RTX PRO 6000. On CUDA, use NVIDIA driver R580+ and CDI.",
          "setupUrl": "https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/cdi-support.html",
          "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/nemotron-3.5.md",
          "engineVersion": "1.2.9",
          "context": 8192,
          "parallel": 1,
          "settings": {
            "defaults": {
              "context": 8192,
              "parallel": 1,
              "input": "text"
            },
            "contexts": [
              8192,
              16384,
              32768
            ],
            "requests": [
              1,
              2,
              4,
              8
            ],
            "inputs": [
              "text"
            ]
          },
          "command": "(\n  set -e\n  image=localhost/tensorfold:1.2.9\n  podman image exists \"$image\" || \\\n    podman build -t \"$image\" --build-arg TENSORFOLD_VERSION=1.2.9 \\\n      'https://github.com/ashhart/TensorFold.git#6b9de695104b6b8b30c20458da87ac06c9480289'\n\n  podman run --rm --pull=never -v tensorfold-weights:/huggingface:U \\\n    \"$image\" pull TensorFold/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit@d9d758fb83953437f7263256b0d96157e2a348b8\n\n  podman run --rm --pull=never --device nvidia.com/gpu=0 \\\n    -p 127.0.0.1:8080:8080 -v tensorfold-weights:/huggingface:ro \\\n    \"$image\" serve \\\n    \"/huggingface/hub/models--TensorFold--NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit/snapshots/d9d758fb83953437f7263256b0d96157e2a348b8\" \\\n    --name nemotron-3.5-lightning --host 0.0.0.0 --port 8080 \\\n    --context 8192 --parallel 1\n)"
        },
        "intel-arc-b70": {
          "format": "MLX 4-bit · MTP",
          "estimateGB": 32,
          "weightGB": 18.5,
          "checkpoint": "https://huggingface.co/TensorFold/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit/tree/d9d758fb83953437f7263256b0d96157e2a348b8",
          "runtime": "Native Intel XPU",
          "setup": "Experimental Intel build. Use the Linux driver and Level Zero setup in the build guide. This checkpoint supports text prompts.",
          "setupUrl": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/xpu/README.md#building",
          "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/xpu-arc.md",
          "engineVersion": "1.2.9",
          "sourceTest": "Intel Arc Pro B70",
          "context": 8192,
          "parallel": 1,
          "settings": {
            "defaults": {
              "context": 8192,
              "parallel": 1,
              "input": "text"
            },
            "contexts": [
              8192,
              16384,
              32768
            ],
            "requests": [
              1,
              2,
              4
            ],
            "inputs": [
              "text"
            ]
          },
          "command": "(\n  set -e\n  export HF_HUB_CACHE=\"${HF_HUB_CACHE:-${HF_HOME:-$HOME/.cache/huggingface}/hub}\"\n  ./zig-out/xpu/bin/tensorfold-native pull TensorFold/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit@d9d758fb83953437f7263256b0d96157e2a348b8\n\n  ./zig-out/xpu/bin/tensorfold-native serve \\\n    \"$HF_HUB_CACHE/models--TensorFold--NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit/snapshots/d9d758fb83953437f7263256b0d96157e2a348b8\" \\\n    --name nemotron-3.5-lightning --host 127.0.0.1 --port 8080 \\\n    --context 8192 --parallel 1\n)",
          "decode": {
            "rate": "157.3–171.8",
            "hardware": "Intel Arc Pro B70",
            "conditions": "One greedy reply with MTP, on the MLX 4-bit checkpoint. Measured across code, prose, lists and copying prompts.",
            "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/xpu-arc.md#measurements",
            "credit": "TensorFold measurements"
          }
        },
        "cuda": {
          "format": "MLX 4-bit · MTP",
          "estimateGB": 28,
          "weightGB": 18.5,
          "checkpoint": "https://huggingface.co/TensorFold/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit/tree/d9d758fb83953437f7263256b0d96157e2a348b8",
          "runtime": "Podman · CUDA",
          "setup": "The engine docs cover Apple Silicon, DGX Spark, RTX 3090 and RTX PRO 6000. On CUDA, use NVIDIA driver R580+ and CDI.",
          "setupUrl": "https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/cdi-support.html",
          "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/nemotron-3.5.md",
          "engineVersion": "1.2.9",
          "context": 8192,
          "parallel": 1,
          "settings": {
            "defaults": {
              "context": 8192,
              "parallel": 1,
              "input": "text"
            },
            "contexts": [
              8192,
              16384,
              32768
            ],
            "requests": [
              1,
              2,
              4,
              8
            ],
            "inputs": [
              "text"
            ]
          },
          "command": "(\n  set -e\n  image=localhost/tensorfold:1.2.9\n  podman image exists \"$image\" || \\\n    podman build -t \"$image\" --build-arg TENSORFOLD_VERSION=1.2.9 \\\n      'https://github.com/ashhart/TensorFold.git#6b9de695104b6b8b30c20458da87ac06c9480289'\n\n  podman run --rm --pull=never -v tensorfold-weights:/huggingface:U \\\n    \"$image\" pull TensorFold/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit@d9d758fb83953437f7263256b0d96157e2a348b8\n\n  podman run --rm --pull=never --device nvidia.com/gpu=0 \\\n    -p 127.0.0.1:8080:8080 -v tensorfold-weights:/huggingface:ro \\\n    \"$image\" serve \\\n    \"/huggingface/hub/models--TensorFold--NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit/snapshots/d9d758fb83953437f7263256b0d96157e2a348b8\" \\\n    --name nemotron-3.5-lightning --host 0.0.0.0 --port 8080 \\\n    --context 8192 --parallel 1\n)"
        }
      },
      "capabilities": {
        "use": "Start with document questions and local chat for several users.",
        "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/nemotron-3.5.md",
        "epoch": null,
        "size": "30B",
        "sizeLabel": "30 billion parameters",
        "sizeSource": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16",
        "shortUse": "Chat for several users",
        "sizeDetails": "Uses 3B parameters at a time while generating text."
      }
    },
    {
      "id": "gemma-4-e4b-intel",
      "name": "Gemma 4 E4B · Nemotron Bolt",
      "provider": "NVIDIA",
      "mark": "nvidia-color.svg",
      "description": "NVIDIA’s Nemotron Bolt E4B, based on Gemma 4, served as text on Intel Arc.",
      "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/gemma-4-native.md",
      "test": "curl -fsS http://127.0.0.1:8080/v1/chat/completions \\\n  -H 'Content-Type: application/json' \\\n  -d '{\"model\":\"gemma-4-e4b-intel\",\"messages\":[{\"role\":\"user\",\"content\":\"Say hello in one sentence.\"}],\"max_tokens\":64}'",
      "deployments": {
        "intel-arc-b70": {
          "format": "Converted MLX 4-bit",
          "estimateGB": null,
          "hardwareGB": 32,
          "weightGB": 16,
          "checkpoint": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-BF16/tree/49843eed8db1f1c616c47af75cfd36e07aec0667",
          "runtime": "Native Intel XPU",
          "setup": "Build the Linux Intel engine, then follow the tested 4-bit conversion. The download size shown is for NVIDIA’s original weights.",
          "setupUrl": "https://github.com/ashhart/TensorFold/pull/1552",
          "source": "https://github.com/ashhart/TensorFold/pull/1552",
          "engineVersion": "1.2.9",
          "sourceTest": "Intel Arc Pro B70",
          "guideLabel": "Open model conversion guide ↗",
          "command": null,
          "decode": {
            "rate": "135.6",
            "hardware": "Intel Arc Pro B70",
            "conditions": "One stream on the converted 4-bit checkpoint in TensorFold 1.2.0. The per-layer input linears kept BF16 precision.",
            "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/CHANGELOG.md#120",
            "credit": "TensorFold release measurements"
          }
        }
      },
      "capabilities": {
        "use": "Start with local text chat, short summaries, and instruction following.",
        "source": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-BF16",
        "epoch": null,
        "size": "8B",
        "sizeLabel": "8 billion parameters",
        "sizeSource": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-BF16",
        "shortUse": "Chat and instructions",
        "sizeDetails": "NVIDIA reports 4B effective parameters for this Gemma 4 model."
      }
    },
    {
      "id": "lfm3-1.7b",
      "name": "LFM3‑1.7B",
      "provider": "Liquid AI",
      "mark": "liquid.svg",
      "description": "A small text model for Intel Iris Xe laptop graphics.",
      "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/xpu-arc.md",
      "test": "curl -fsS http://127.0.0.1:8080/v1/chat/completions \\\n  -H 'Content-Type: application/json' \\\n  -d '{\"model\":\"lfm3-1.7b\",\"messages\":[{\"role\":\"user\",\"content\":\"Say hello in one sentence.\"}],\"max_tokens\":64}'",
      "deployments": {
        "intel-iris-xe": {
          "format": "MLX 4-bit",
          "estimateGB": 3,
          "weightGB": 1,
          "checkpoint": "https://huggingface.co/LiquidAI/LFM3-1.7B-MLX-4bit/tree/d0ca73dcfa70ce18733754a492cdc1fb0123262d",
          "runtime": "Native Intel XPU",
          "setup": "Build TensorFold 1.2.9 in Ubuntu on WSL2 with zig build xpu-native -Dxpu-device=tgllp. Run the command from that checkout.",
          "setupUrl": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/xpu-arc.md#tested-setup",
          "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/xpu-arc.md#tested-setup",
          "engineVersion": "1.2.9",
          "sourceTest": "Intel Iris Xe · Windows with WSL2",
          "context": 8192,
          "parallel": 1,
          "settings": {
            "defaults": {
              "context": 8192,
              "parallel": 1,
              "input": "text"
            },
            "contexts": [
              8192,
              16384
            ],
            "requests": [
              1,
              2,
              4
            ],
            "inputs": [
              "text"
            ]
          },
          "command": "(\n  set -e\n  export HF_HUB_CACHE=\"${HF_HUB_CACHE:-${HF_HOME:-$HOME/.cache/huggingface}/hub}\"\n  ./zig-out/xpu/bin/tensorfold-native pull LiquidAI/LFM3-1.7B-MLX-4bit@d0ca73dcfa70ce18733754a492cdc1fb0123262d\n\n  ./zig-out/xpu/bin/tensorfold-native serve \\\n    \"$HF_HUB_CACHE/models--LiquidAI--LFM3-1.7B-MLX-4bit/snapshots/d0ca73dcfa70ce18733754a492cdc1fb0123262d\" \\\n    --name lfm3-1.7b --host 127.0.0.1 --port 8080 \\\n    --context 8192 --parallel 1\n)",
          "decode": {
            "rate": "15.9–26.3",
            "hardware": "Intel Iris Xe · WSL2",
            "conditions": "Greedy chat and prose, one request with copy drafts. Tested in Ubuntu 24.04 under WSL2 on Gen12LP.",
            "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/xpu-arc.md#tested-setup",
            "credit": "TensorFold measurements"
          }
        }
      },
      "capabilities": {
        "use": "Start with short questions, simple rewrites, and extracting information from text.",
        "source": "https://huggingface.co/LiquidAI/LFM3-1.7B-MLX-4bit",
        "epoch": null,
        "size": "1.7B",
        "sizeLabel": "1.7 billion parameters",
        "sizeSource": "https://huggingface.co/LiquidAI/LFM3-1.7B-MLX-4bit",
        "shortUse": "Short text tasks"
      }
    },
    {
      "id": "qwen3.6-35b-a3b",
      "name": "Qwen3.6-35B-A3B",
      "provider": "Qwen",
      "mark": "qwen-color.svg",
      "description": "A mixture-of-experts model with 3B active parameters and a built-in MTP head.",
      "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/qwen3.6-moe.md",
      "test": "curl -fsS http://127.0.0.1:8080/v1/chat/completions \\\n  -H 'Content-Type: application/json' \\\n  -d '{\"model\":\"qwen3.6-35b-a3b\",\"messages\":[{\"role\":\"user\",\"content\":\"Say hello in one sentence.\"}],\"max_tokens\":64}'",
      "deployments": {
        "dgx-spark": {
          "format": "NVFP4 · MTP",
          "estimateGB": 32,
          "weightGB": 23.4,
          "checkpoint": "https://huggingface.co/nvidia/Qwen3.6-35B-A3B-NVFP4/tree/1355db6a052410cfd62085d94b58866fd0f2c3c5",
          "runtime": "Podman · CUDA",
          "setup": "For one DGX Spark with Podman, NVIDIA driver R580+ and CDI. The launcher builds TensorFold and uses the checkpoint’s MTP head.",
          "setupUrl": "https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/cdi-support.html",
          "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/qwen3.6-moe.md#nvidias-nvfp4-checkpoint-native-cuda",
          "engineVersion": "1.2.9",
          "context": 8192,
          "parallel": 1,
          "settings": {
            "defaults": {
              "context": 8192,
              "parallel": 1,
              "input": "text"
            },
            "contexts": [
              8192,
              16384,
              32768,
              65536
            ],
            "requests": [
              1,
              2
            ],
            "inputs": [
              "text"
            ]
          },
          "command": "(\n  set -e\n  image=localhost/tensorfold:1.2.9\n  podman image exists \"$image\" || \\\n    podman build -t \"$image\" --build-arg TENSORFOLD_VERSION=1.2.9 \\\n      'https://github.com/ashhart/TensorFold.git#6b9de695104b6b8b30c20458da87ac06c9480289'\n\n  podman run --rm --pull=never -v tensorfold-weights:/huggingface:U \\\n    \"$image\" pull nvidia/Qwen3.6-35B-A3B-NVFP4@1355db6a052410cfd62085d94b58866fd0f2c3c5\n\n  podman run --rm --pull=never --device nvidia.com/gpu=0 \\\n    -p 127.0.0.1:8080:8080 -v tensorfold-weights:/huggingface:ro \\\n    \"$image\" serve \\\n    \"/huggingface/hub/models--nvidia--Qwen3.6-35B-A3B-NVFP4/snapshots/1355db6a052410cfd62085d94b58866fd0f2c3c5\" \\\n    --drafter \"/huggingface/hub/models--nvidia--Qwen3.6-35B-A3B-NVFP4/snapshots/1355db6a052410cfd62085d94b58866fd0f2c3c5\" \\\n    --name qwen3.6-35b-a3b --host 0.0.0.0 --port 8080 \\\n    --context 8192 --parallel 1\n)",
          "decode": {
            "rate": "111.8",
            "hardware": "One DGX Spark",
            "conditions": "One greedy chat reply with thinking off, NVFP4 and MTP.",
            "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/qwen3.6-moe.md#nvidias-nvfp4-checkpoint-native-cuda",
            "credit": "TensorFold measurements"
          }
        }
      },
      "codingReference": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B",
      "capabilities": {
        "use": "Start with code explanation, patches with tests and tool-assisted text workflows.",
        "source": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B",
        "epoch": {
          "name": "Qwen 3.6 35B-A3B",
          "score": 144.11,
          "low": 142.23,
          "high": 146
        },
        "size": "35B",
        "sizeLabel": "35 billion parameters",
        "sizeSource": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B",
        "shortUse": "Coding and tools",
        "sizeDetails": "Uses 3B parameters at a time while generating text."
      }
    },
    {
      "id": "spark-x2.6-4b",
      "name": "Spark-X2.6-4B",
      "provider": "XHToken",
      "description": "A compact BF16 text model with a documented native Metal build.",
      "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/spark-x2.6-4b-metal.md",
      "test": "curl -fsS http://127.0.0.1:8080/v1/chat/completions \\\n  -H 'Content-Type: application/json' \\\n  -d '{\"model\":\"spark-x2.6-4b\",\"messages\":[{\"role\":\"user\",\"content\":\"Say hello in one sentence.\"}],\"max_tokens\":64}'",
      "deployments": {
        "apple-silicon": {
          "format": "BF16",
          "estimateGB": null,
          "hardwareGB": 256,
          "weightGB": 8.2,
          "checkpoint": "https://huggingface.co/XHToken/Spark-X2.6-4B/tree/66e3ae7bee54e0a523d692cb8ef841f35409807b",
          "runtime": "Native Metal",
          "setup": "The documented setup uses one M5 Ultra with 256 GB unified memory. Follow the guide to build the native engine and serve the pinned BF16 checkpoint.",
          "setupUrl": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/spark-x2.6-4b-metal.md#build-and-serve",
          "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/spark-x2.6-4b-metal.md#build-and-serve",
          "engineVersion": "1.2.9",
          "command": null,
          "decode": {
            "rate": "106",
            "hardware": "M5 Ultra",
            "conditions": "One decode stream with the original BF16 checkpoint.",
            "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/spark-x2.6-4b-metal.md",
            "credit": "TensorFold measurements"
          }
        }
      },
      "capabilities": {
        "use": "Start with short local chat, text tagging and small batch jobs.",
        "source": "https://huggingface.co/XHToken/Spark-X2.6-4B",
        "epoch": null,
        "size": "4B",
        "sizeLabel": "4 billion parameters",
        "sizeSource": "https://huggingface.co/XHToken/Spark-X2.6-4B",
        "shortUse": "Short chat and tagging"
      }
    },
    {
      "id": "nemotron-3.5-super-vl",
      "name": "Nemotron 3.5 Super VL",
      "provider": "NVIDIA",
      "mark": "nvidia-color.svg",
      "description": "A 120B mixture-of-experts model with optional image input.",
      "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/nemotron-3.5-super-vl.md",
      "test": "curl -fsS http://127.0.0.1:8080/v1/chat/completions \\\n  -H 'Content-Type: application/json' \\\n  -d '{\"model\":\"nemotron-3.5-super-vl\",\"messages\":[{\"role\":\"user\",\"content\":\"Say hello in one sentence.\"}],\"max_tokens\":64}'",
      "deployments": {
        "dgx-spark": {
          "format": "NVFP4",
          "estimateGB": 96,
          "weightGB": 75,
          "checkpoint": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4/tree/82d93699acd3be9e4617c53360fe5b3f5dbeb1b3",
          "runtime": "Podman · CUDA",
          "setup": "Requires Podman, NVIDIA driver R580+ and NVIDIA CDI. The first run builds the image and caches the weights.",
          "setupUrl": "https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/cdi-support.html",
          "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/nemotron-3.5-super-vl.md",
          "engineVersion": "1.2.9",
          "context": 8192,
          "parallel": 1,
          "settings": {
            "defaults": {
              "context": 8192,
              "parallel": 1,
              "input": "text"
            },
            "contexts": [
              8192,
              16384,
              32768
            ],
            "requests": [
              1,
              2,
              4
            ],
            "inputs": [
              "text",
              "images"
            ]
          },
          "command": "(\n  set -e\n  image=localhost/tensorfold:1.2.9\n  podman image exists \"$image\" || \\\n    podman build -t \"$image\" --build-arg TENSORFOLD_VERSION=1.2.9 \\\n      'https://github.com/ashhart/TensorFold.git#6b9de695104b6b8b30c20458da87ac06c9480289'\n\n  podman run --rm --pull=never -v tensorfold-weights:/huggingface:U \\\n    \"$image\" pull nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4@82d93699acd3be9e4617c53360fe5b3f5dbeb1b3\n\n  podman run --rm --pull=never --device nvidia.com/gpu=0 \\\n    -p 127.0.0.1:8080:8080 -v tensorfold-weights:/huggingface:ro \\\n    \"$image\" serve \\\n    \"/huggingface/hub/models--nvidia--NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4/snapshots/82d93699acd3be9e4617c53360fe5b3f5dbeb1b3\" \\\n    --name nemotron-3.5-super-vl --host 0.0.0.0 --port 8080 \\\n    --context 8192 --parallel 1\n)"
        }
      },
      "capabilities": {
        "use": "Start with questions about screenshots and charts. Choose image input in Launch settings.",
        "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/nemotron-3.5-super-vl.md#images",
        "epoch": null,
        "size": "120B",
        "sizeLabel": "120 billion parameters",
        "sizeSource": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4",
        "shortUse": "Screenshots and charts",
        "sizeDetails": "Uses 12B parameters at a time while generating text."
      }
    },
    {
      "id": "qwen3.8-flash-next",
      "name": "Qwen3.8 Flash Next",
      "provider": "Qwen",
      "mark": "qwen-color.svg",
      "description": "A sparse mixture-of-experts model with an embedded MTP draft head.",
      "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/qwen3.8-flash-next.md",
      "test": "curl -fsS http://127.0.0.1:8080/v1/chat/completions \\\n  -H 'Content-Type: application/json' \\\n  -d '{\"model\":\"qwen3.8-flash-next\",\"messages\":[{\"role\":\"user\",\"content\":\"Say hello in one sentence.\"}],\"max_tokens\":64}'",
      "deployments": {
        "apple-silicon": {
          "format": "MLX 6-bit",
          "estimateGB": 256,
          "weightGB": 158.1,
          "checkpoint": "https://huggingface.co/TensorFold/Qwen3.8-Flash-Next-MLX-6bit-MTP/tree/45032a193a5d4126ba8534822cbaf8fb30b3e368",
          "runtime": "Native Metal",
          "setup": "The native recipe is documented on an M5 Ultra. Allow space for the weights and the locally prepared model pack.",
          "setupUrl": "https://github.com/ashhart/TensorFold/blob/v1.2.9/README.md#mac-with-apple-silicon",
          "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/qwen3.8-flash-next.md",
          "engineVersion": "1.2.9",
          "context": 8192,
          "parallel": 1,
          "settings": {
            "defaults": {
              "context": 8192,
              "parallel": 1,
              "input": "text"
            },
            "contexts": [
              8192,
              16384,
              32768,
              65536,
              131072
            ],
            "requests": [
              1,
              2,
              4,
              8
            ],
            "inputs": [
              "text",
              "video"
            ]
          },
          "command": "(\n  set -e\n  export HF_HUB_CACHE=\"${HF_HUB_CACHE:-${HF_HOME:-$HOME/.cache/huggingface}/hub}\"\n  tensorfold pull TensorFold/Qwen3.8-Flash-Next-MLX-6bit-MTP@45032a193a5d4126ba8534822cbaf8fb30b3e368\n\n  tensorfold serve \\\n    \"$HF_HUB_CACHE/models--TensorFold--Qwen3.8-Flash-Next-MLX-6bit-MTP/snapshots/45032a193a5d4126ba8534822cbaf8fb30b3e368\" \\\n    --name qwen3.8-flash-next --host 127.0.0.1 --port 8080 \\\n    --context 8192 --parallel 1\n)",
          "decode": {
            "rate": "137–216",
            "hardware": "M5 Ultra",
            "conditions": "One stream at a time, MLX 6-bit MTP, with a 1M-token session and vision enabled.",
            "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/qwen3.8-flash-next.md",
            "credit": "TensorFold measurements"
          }
        },
        "dgx-spark": {
          "format": "NVIDIA FP4 QAD",
          "estimateGB": 120,
          "weightGB": 102.2,
          "checkpoint": "https://huggingface.co/nvidia/Qwen3.8-Flash-Next-FP4-Local/tree/b9ce18a6887e18c089fdea8509a33e63429a3300",
          "runtime": "Podman",
          "setup": "MiaAI-Lab’s single-Spark image, TensorFold 1.1.4. Podman needs working NVIDIA CDI. The launcher uses BF16 KV cache.",
          "setupUrl": "https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/cdi-support.html",
          "source": "https://github.com/MiaAI-Lab/Qwen3.8-Flash-Next-Single-DGX-Spark-TensorFold/tree/20e15e41f2f514a0d287710f4665ee1b5009ba41",
          "engineVersion": "1.1.4",
          "creator": {
            "name": "Mia’s AI Lab",
            "website": "https://mia-ai.net/",
            "logo": "/assets/mia-ai-lab-logo.png",
            "article": "https://mia-ai.net/models/Qwen3.8-Flash-Next-Single-DGX-Spark-TensorFold",
            "podman": true
          },
          "image": "ghcr.io/miaai-lab/qwen3.8-flash-next-single-dgx-spark-tensorfold@sha256:5c6c42b9747318c1cd8b6cb85cb69493fff2e82cb23b61e49b9cf54cd09ac28e",
          "context": 8192,
          "parallel": 1,
          "settings": {
            "defaults": {
              "context": 8192,
              "parallel": 1,
              "input": "text"
            },
            "contexts": [
              8192,
              16384,
              32768,
              65536
            ],
            "requests": [
              1,
              2,
              4
            ],
            "inputs": [
              "text",
              "video"
            ]
          },
          "command": "(\n  set -e\n  image='ghcr.io/miaai-lab/qwen3.8-flash-next-single-dgx-spark-tensorfold@sha256:5c6c42b9747318c1cd8b6cb85cb69493fff2e82cb23b61e49b9cf54cd09ac28e'\n  podman image exists \"$image\" || podman pull \"$image\"\n\n  podman run --rm -i --pull=never --entrypoint python \\\n    -v tensorfold-mia-weights:/root/.cache/huggingface \"$image\" - \\\n    nvidia/Qwen3.8-Flash-Next-FP4-Local@b9ce18a6887e18c089fdea8509a33e63429a3300 <<'PY'\nimport json\nimport sys\nfrom pathlib import Path\nfrom tempfile import NamedTemporaryFile\nfrom huggingface_hub import snapshot_download\n\ncache = Path(\"/root/.cache/huggingface/hub\")\nfor artifact in sys.argv[1:]:\n    repo, revision = artifact.rsplit(\"@\", 1)\n    snapshot = cache / (\"models--\" + repo.replace(\"/\", \"--\")) / \"snapshots\" / revision\n    receipt = snapshot / \".tensorfold-recipe-files.json\"\n    try:\n        files = json.loads(receipt.read_text())\n        if isinstance(files, dict) and files and all((snapshot / name).stat().st_size == size for name, size in files.items()):\n            print(\"Reusing cached snapshot \" + artifact, flush=True)\n            continue\n    except (OSError, ValueError):\n        pass\n    print(\"Downloading \" + artifact, flush=True)\n    downloaded = Path(snapshot_download(repo_id=repo, revision=revision, cache_dir=str(cache)))\n    files = {p.relative_to(downloaded).as_posix(): p.stat().st_size\n             for p in downloaded.rglob(\"*\") if p.is_file() and not p.name.startswith(\".tensorfold-recipe-\")}\n    if \"config.json\" not in files or not any(name.endswith(\".safetensors\") for name in files):\n        raise RuntimeError(\"Incomplete snapshot \" + artifact)\n    with NamedTemporaryFile(mode=\"w\", dir=downloaded, prefix=\".tensorfold-recipe-\", delete=False) as temp:\n        json.dump(files, temp)\n    Path(temp.name).replace(receipt)\nPY\n\n  podman run --rm --pull=never --device nvidia.com/gpu=0 \\\n    -p 127.0.0.1:8080:8080 -v tensorfold-mia-weights:/root/.cache/huggingface:ro \\\n    -v tensorfold-mia-flash-next-kernels:/cache --shm-size=1g \\\n    -e HF_HUB_OFFLINE=1 -e TENSORFOLD_NO_UPDATE_CHECK=1 \\\n    -e TENSORFOLD_KV_DTYPE=bf16 -e TENSORFOLD_MEMORY_RESERVE_GIB=9 \\\n    \"$image\" tensorfold-native serve \\\n    /root/.cache/huggingface/hub/models--nvidia--Qwen3.8-Flash-Next-FP4-Local/snapshots/b9ce18a6887e18c089fdea8509a33e63429a3300 \\\n    --kv-cache bf16 --no-thinking --name qwen3.8-flash-next \\\n    --host 0.0.0.0 --port 8080 --context 8192 --parallel 1\n)",
          "decode": {
            "rate": "55.6",
            "hardware": "One DGX Spark",
            "conditions": "One prose reply, NVIDIA FP4 QAD. Mia’s benchmark uses FP8 KV. The launcher here uses BF16 KV.",
            "source": "https://github.com/MiaAI-Lab/Qwen3.8-Flash-Next-Single-DGX-Spark-TensorFold/blob/20e15e41f2f514a0d287710f4665ee1b5009ba41/README.md#performance",
            "credit": "Mia’s AI Lab"
          }
        },
        "dgx-spark-pair": {
          "format": "NVFP4",
          "estimateGB": null,
          "weightGB": 132.7,
          "checkpoint": "https://huggingface.co/nvidia/Qwen3.8-Flash-Next-NVFP4/tree/fc694b54fb0174e0913e6adf86691ef85a4ead47",
          "runtime": "Native CUDA",
          "setup": "Experimental native CUDA setup. Use the guide to configure NCCL on both Sparks and start rank 1 first.",
          "setupUrl": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/flash-next-cuda.md#nvidias-nvfp4-checkpoint-on-two-gb10s",
          "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/flash-next-cuda.md#nvidias-nvfp4-checkpoint-on-two-gb10s",
          "engineVersion": "1.2.9",
          "command": null
        },
        "amd-hip": {
          "format": "oQ4e",
          "estimateGB": 128,
          "weightGB": 115.1,
          "checkpoint": "https://huggingface.co/unigilby/Qwen3.8-Flash-Next-oQ4e-HIP/tree/5a0fb4147f6f7e81bbd5ec2d8a2a3ee3f648427d",
          "runtime": "Native HIP",
          "setup": "For Radeon 8060S (gfx1151) with 128 GB unified memory. Build the HIP binary from the setup guide, then run this from the checkout.",
          "setupUrl": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/qwen3.8-flash-next.md#native-hip-on-a-strix-halo-igpu",
          "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/qwen3.8-flash-next.md#native-hip-on-a-strix-halo-igpu",
          "engineVersion": "1.2.9",
          "context": 8192,
          "parallel": 1,
          "settings": {
            "defaults": {
              "context": 8192,
              "parallel": 1,
              "input": "text"
            },
            "contexts": [
              8192,
              16384,
              32768,
              65536,
              131072
            ],
            "requests": [
              1
            ],
            "inputs": [
              "text"
            ]
          },
          "command": "(\n  set -e\n  export HF_HUB_CACHE=\"${HF_HUB_CACHE:-${HF_HOME:-$HOME/.cache/huggingface}/hub}\"\n  ./zig-out/native-hip/bin/tensorfold-native pull unigilby/Qwen3.8-Flash-Next-oQ4e-HIP@5a0fb4147f6f7e81bbd5ec2d8a2a3ee3f648427d\n\n  ./zig-out/native-hip/bin/tensorfold-native serve \\\n    \"$HF_HUB_CACHE/models--unigilby--Qwen3.8-Flash-Next-oQ4e-HIP/snapshots/5a0fb4147f6f7e81bbd5ec2d8a2a3ee3f648427d\" \\\n    --name qwen3.8-flash-next --host 127.0.0.1 --port 8080 \\\n    --context 8192 --parallel 1\n)",
          "taskCheck": {
            "revision": "5a0fb4147f6f7e81bbd5ec2d8a2a3ee3f648427d",
            "summary": "93.9% HumanEval pass@1 and 69.1% MMLU-Pro on Radeon 8060S with this oQ4e checkpoint. Greedy decoding, thinking off. MMLU-Pro used a 700-question sample.",
            "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/qwen3.8-flash-next.md#native-hip-on-a-strix-halo-igpu"
          },
          "decode": {
            "rate": "36.2",
            "hardware": "Strix Halo · Radeon 8060S",
            "conditions": "One chat stream, oQ4e checkpoint, 131K context and three MTP drafts.",
            "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/qwen3.8-flash-next.md#native-hip-on-a-strix-halo-igpu",
            "credit": "TensorFold measurements"
          }
        }
      },
      "codingReference": "https://huggingface.co/Qwen/Qwen3.8-Flash-Next#benchmark-results",
      "capabilities": {
        "use": "Start with multi-file coding and long documents. Check the device recipe for image and video support.",
        "source": "https://huggingface.co/Qwen/Qwen3.8-Flash-Next#benchmark-results",
        "epoch": null,
        "size": "125B",
        "sizeLabel": "125 billion parameters in the main network",
        "sizeSource": "https://huggingface.co/Qwen/Qwen3.8-Flash-Next#benchmark-results",
        "shortUse": "Coding and long documents",
        "sizeDetails": "Qwen reports 6B active parameters, plus 51B for its text lookup table and 4B for predicting ahead.",
        "referenceCheck": "Qwen reports 62.5% on its corrected SWE-bench Pro task set using Claude Code at 256K context."
      }
    },
    {
      "id": "qwen3.8-27b",
      "name": "Qwen3.8-27B",
      "provider": "Qwen",
      "mark": "qwen-color.svg",
      "description": "Dense reasoning model with a DFlash2 drafter.",
      "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/qwen3.8-27b.md",
      "test": "curl -fsS http://127.0.0.1:8080/v1/chat/completions \\\n  -H 'Content-Type: application/json' \\\n  -d '{\"model\":\"qwen3.8-27b\",\"messages\":[{\"role\":\"user\",\"content\":\"Say hello in one sentence.\"}],\"max_tokens\":64}'",
      "deployments": {
        "apple-silicon": {
          "format": "MLX 4-bit",
          "estimateGB": 40,
          "weightGB": 19.9,
          "checkpoint": "https://huggingface.co/TensorFold/Qwen3.8-27B-MLX-4bit/tree/22d8d538154e0e5b6f8dcb1fcc74a60a0798104c",
          "runtime": "Native Metal",
          "setup": "TensorFold CLI on an Apple Silicon Mac. The engine recipe documents M5 Max and M3 Ultra runs.",
          "setupUrl": "https://github.com/ashhart/TensorFold/blob/v1.2.9/README.md#mac-with-apple-silicon",
          "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/qwen3.8-27b.md",
          "engineVersion": "1.2.9",
          "context": 8192,
          "parallel": 1,
          "settings": {
            "defaults": {
              "context": 8192,
              "parallel": 1,
              "input": "text"
            },
            "contexts": [
              8192,
              16384,
              32768,
              65536
            ],
            "requests": [
              1,
              2,
              4,
              8
            ],
            "inputs": [
              "text",
              "images"
            ]
          },
          "command": "(\n  set -e\n  export HF_HUB_CACHE=\"${HF_HUB_CACHE:-${HF_HOME:-$HOME/.cache/huggingface}/hub}\"\n  tensorfold pull TensorFold/Qwen3.8-27B-MLX-4bit@22d8d538154e0e5b6f8dcb1fcc74a60a0798104c\n  tensorfold pull z-lab/Qwen3.8-27B-DFlash2@50307d4c4cde6860d4eee73e2547cd786fe8e8a4\n\n  tensorfold serve \\\n    \"$HF_HUB_CACHE/models--TensorFold--Qwen3.8-27B-MLX-4bit/snapshots/22d8d538154e0e5b6f8dcb1fcc74a60a0798104c\" \\\n    --drafter \"$HF_HUB_CACHE/models--z-lab--Qwen3.8-27B-DFlash2/snapshots/50307d4c4cde6860d4eee73e2547cd786fe8e8a4\" \\\n    --name qwen3.8-27b --host 127.0.0.1 --port 8080 \\\n    --context 8192 --parallel 1\n)",
          "decode": {
            "rate": "102",
            "hardware": "M5 Ultra",
            "conditions": "One greedy reply after an 8K prompt, MLX 4-bit with DFlash2.",
            "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/qwen3.8-27b.md",
            "credit": "TensorFold measurements"
          }
        },
        "dgx-spark": {
          "format": "NVFP4",
          "estimateGB": 40,
          "weightGB": 25.8,
          "checkpoint": "https://huggingface.co/nvidia/Qwen3.8-27B-NVFP4/tree/482ca0f3832238542f8f5295dde86b5f22711d80",
          "runtime": "Podman",
          "setup": "Podman on Linux, NVIDIA driver R580 or newer and NVIDIA CDI. Podman 4.x needs a CDI 0.6-compatible spec. First run pulls MiaAI-Lab’s image (TensorFold 0.6.0) and caches the pinned weights.",
          "setupUrl": "https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/cdi-support.html",
          "source": "https://github.com/MiaAI-Lab/Qwen3.8-27B-DGX-Spark-TensorFold/tree/c11ff138871cf43d12898d79dbfa4dc952d17ca4",
          "engineVersion": "0.6.0",
          "creator": {
            "name": "Mia’s AI Lab",
            "website": "https://mia-ai.net/",
            "logo": "/assets/mia-ai-lab-logo.png",
            "article": "https://mia-ai.net/models/Qwen3.8-27B-DGX-Spark-TensorFold",
            "podman": true
          },
          "image": "ghcr.io/miaai-lab/qwen3.8-27b-dgx-spark-tensorfold@sha256:043d192b9e1725f55c5439fc9672055de891433e5b31448ed8963918e3c460e9",
          "smokeTest": {
            "commandSha256": "066971dda900ec1d112579226c822eb7e1bcfa578573a7bffb7c68600fb222e5",
            "hardware": "DGX Spark",
            "date": "2026-11-18",
            "coldReplySeconds": 541,
            "warmReplySeconds": 18,
            "reusedWeights": true
          },
          "context": 8192,
          "parallel": 1,
          "settings": {
            "defaults": {
              "context": 8192,
              "parallel": 1,
              "input": "text"
            },
            "contexts": [
              8192,
              16384,
              32768,
              65536
            ],
            "requests": [
              1,
              2,
              4
            ],
            "inputs": [
              "text"
            ]
          },
          "command": "(\n  set -e\n  image='ghcr.io/miaai-lab/qwen3.8-27b-dgx-spark-tensorfold@sha256:043d192b9e1725f55c5439fc9672055de891433e5b31448ed8963918e3c460e9'\n  podman image exists \"$image\" || podman pull \"$image\"\n\n  podman run --rm -i --pull=never --entrypoint python \\\n    -v tensorfold-mia-weights:/root/.cache/huggingface \"$image\" - \\\n    nvidia/Qwen3.8-27B-NVFP4@482ca0f3832238542f8f5295dde86b5f22711d80 z-lab/Qwen3.8-27B-DFlash2@50307d4c4cde6860d4eee73e2547cd786fe8e8a4 <<'PY'\nimport json\nimport sys\nfrom pathlib import Path\nfrom tempfile import NamedTemporaryFile\nfrom huggingface_hub import snapshot_download\n\ncache = Path(\"/root/.cache/huggingface/hub\")\nfor artifact in sys.argv[1:]:\n    repo, revision = artifact.rsplit(\"@\", 1)\n    snapshot = cache / (\"models--\" + repo.replace(\"/\", \"--\")) / \"snapshots\" / revision\n    receipt = snapshot / \".tensorfold-recipe-files.json\"\n    try:\n        files = json.loads(receipt.read_text())\n        if isinstance(files, dict) and files and all((snapshot / name).stat().st_size == size for name, size in files.items()):\n            print(\"Reusing cached snapshot \" + artifact, flush=True)\n            continue\n    except (OSError, ValueError):\n        pass\n    print(\"Downloading \" + artifact, flush=True)\n    downloaded = Path(snapshot_download(repo_id=repo, revision=revision, cache_dir=str(cache)))\n    files = {p.relative_to(downloaded).as_posix(): p.stat().st_size\n             for p in downloaded.rglob(\"*\") if p.is_file() and not p.name.startswith(\".tensorfold-recipe-\")}\n    if \"config.json\" not in files or not any(name.endswith(\".safetensors\") for name in files):\n        raise RuntimeError(\"Incomplete snapshot \" + artifact)\n    with NamedTemporaryFile(mode=\"w\", dir=downloaded, prefix=\".tensorfold-recipe-\", delete=False) as temp:\n        json.dump(files, temp)\n    Path(temp.name).replace(receipt)\nPY\n\n  podman run --rm --pull=never --device nvidia.com/gpu=0 \\\n    -p 127.0.0.1:8080:8080 -v tensorfold-mia-weights:/root/.cache/huggingface:ro \\\n    -v tensorfold-mia-kernels:/cache --shm-size=1g \\\n    -e HF_HUB_OFFLINE=1 -e TENSORFOLD_NO_UPDATE_CHECK=1 \\\n    -e TENSORFOLD_KV_DTYPE=bf16 -e TENSORFOLD_MEMORY_RESERVE_GIB=8 \\\n    \"$image\" tensorfold serve \\\n    /root/.cache/huggingface/hub/models--nvidia--Qwen3.8-27B-NVFP4/snapshots/482ca0f3832238542f8f5295dde86b5f22711d80 \\\n    --drafter /root/.cache/huggingface/hub/models--z-lab--Qwen3.8-27B-DFlash2/snapshots/50307d4c4cde6860d4eee73e2547cd786fe8e8a4 \\\n    --no-prefill-fp8 --no-thinking --name qwen3.8-27b \\\n    --host 0.0.0.0 --port 8080 --context 8192 --parallel 1\n)",
          "decode": {
            "rate": "38.2",
            "hardware": "One DGX Spark",
            "conditions": "Sampled chat on the same NVIDIA NVFP4 checkpoint in the earlier Python engine. The Podman launch test measured startup and cache reuse.",
            "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/qwen3.8-27b.md#nvfp4-checkpoints",
            "credit": "TensorFold reference measurement"
          }
        },
        "intel-arc-b70": {
          "format": "GPTQ 4-bit · DFlash2",
          "estimateGB": 32,
          "weightGB": 23.4,
          "checkpoint": "https://huggingface.co/SergiioB/Qwen3.8-27B-GPTQ-Int4-sym-G128-MTP-BF16/tree/9d189a60e4c0ad7f9f47cd94bfa393ca10b3924e",
          "runtime": "Native Intel XPU",
          "setup": "Build TensorFold 1.2.9 for Intel in the source checkout. The build guide covers the Linux driver, Level Zero, ocloc and oneDNN headers. Run the command from that checkout.",
          "setupUrl": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/xpu/README.md#building",
          "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/xpu/RECIPE-qwen38-27b-b70.md",
          "engineVersion": "1.2.9",
          "sourceTest": "Intel Arc Pro B70",
          "context": 8192,
          "parallel": 1,
          "settings": {
            "defaults": {
              "context": 8192,
              "parallel": 1,
              "input": "text"
            },
            "contexts": [
              8192,
              16384
            ],
            "requests": [
              1,
              2,
              4
            ],
            "inputs": [
              "text"
            ]
          },
          "command": "(\n  set -e\n  export HF_HUB_CACHE=\"${HF_HUB_CACHE:-${HF_HOME:-$HOME/.cache/huggingface}/hub}\"\n  ./zig-out/xpu/bin/tensorfold-native pull SergiioB/Qwen3.8-27B-GPTQ-Int4-sym-G128-MTP-BF16@9d189a60e4c0ad7f9f47cd94bfa393ca10b3924e\n  ./zig-out/xpu/bin/tensorfold-native pull z-lab/Qwen3.8-27B-DFlash2@50307d4c4cde6860d4eee73e2547cd786fe8e8a4\n\n  ./zig-out/xpu/bin/tensorfold-native serve \\\n    \"$HF_HUB_CACHE/models--SergiioB--Qwen3.8-27B-GPTQ-Int4-sym-G128-MTP-BF16/snapshots/9d189a60e4c0ad7f9f47cd94bfa393ca10b3924e\" \\\n    --drafter \"$HF_HUB_CACHE/models--z-lab--Qwen3.8-27B-DFlash2/snapshots/50307d4c4cde6860d4eee73e2547cd786fe8e8a4\" \\\n    --name qwen3.8-27b --host 127.0.0.1 --port 8080 \\\n    --context 8192 --parallel 1\n)",
          "decode": {
            "rate": "90–133",
            "hardware": "Intel Arc Pro B70",
            "conditions": "One greedy chat or code reply with GPTQ 4-bit and DFlash2. The source run used a 16,384-token context and 230 W.",
            "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/xpu/README.md",
            "credit": "TensorFold XPU measurements"
          }
        },
        "cuda": {
          "format": "NVFP4",
          "estimateGB": 40,
          "weightGB": 25.8,
          "checkpoint": "https://huggingface.co/nvidia/Qwen3.8-27B-NVFP4/tree/482ca0f3832238542f8f5295dde86b5f22711d80",
          "runtime": "Podman · CUDA",
          "setup": "Requires Podman, NVIDIA driver R580+ and NVIDIA CDI. The first run builds the image and caches the weights.",
          "setupUrl": "https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/cdi-support.html",
          "source": "https://github.com/ashhart/TensorFold/blob/v1.2.9/docs/recipes/qwen3.8-27b.md",
          "engineVersion": "1.2.9",
          "context": 8192,
          "parallel": 1,
          "settings": {
            "defaults": {
              "context": 8192,
              "parallel": 1,
              "input": "text"
            },
            "contexts": [
              8192,
              16384,
              32768
            ],
            "requests": [
              1,
              2,
              4
            ],
            "inputs": [
              "text"
            ]
          },
          "command": "(\n  set -e\n  image=localhost/tensorfold:1.2.9\n  podman image exists \"$image\" || \\\n    podman build -t \"$image\" --build-arg TENSORFOLD_VERSION=1.2.9 \\\n      'https://github.com/ashhart/TensorFold.git#6b9de695104b6b8b30c20458da87ac06c9480289'\n\n  podman run --rm --pull=never -v tensorfold-weights:/huggingface:U \\\n    \"$image\" pull nvidia/Qwen3.8-27B-NVFP4@482ca0f3832238542f8f5295dde86b5f22711d80\n  podman run --rm --pull=never -v tensorfold-weights:/huggingface:U \\\n    \"$image\" pull z-lab/Qwen3.8-27B-DFlash2@50307d4c4cde6860d4eee73e2547cd786fe8e8a4\n\n  podman run --rm --pull=never --device nvidia.com/gpu=0 \\\n    -p 127.0.0.1:8080:8080 -v tensorfold-weights:/huggingface:ro \\\n    \"$image\" serve \\\n    \"/huggingface/hub/models--nvidia--Qwen3.8-27B-NVFP4/snapshots/482ca0f3832238542f8f5295dde86b5f22711d80\" \\\n    --drafter \"/huggingface/hub/models--z-lab--Qwen3.8-27B-DFlash2/snapshots/50307d4c4cde6860d4eee73e2547cd786fe8e8a4\" \\\n    --name qwen3.8-27b --host 0.0.0.0 --port 8080 \\\n    --context 8192 --parallel 1\n)"
        }
      },
      "codingReference": "https://huggingface.co/Qwen/Qwen3.8-27B#benchmark-results",
      "capabilities": {
        "use": "Start with code review, debugging and focused reasoning over documents.",
        "source": "https://huggingface.co/Qwen/Qwen3.8-27B#benchmark-results",
        "epoch": {
          "name": "Qwen 3.8 27B",
          "score": 152.65,
          "low": 150.7,
          "high": 154.76
        },
        "size": "27B",
        "sizeLabel": "27 billion parameters",
        "sizeSource": "https://huggingface.co/Qwen/Qwen3.8-27B#benchmark-results",
        "shortUse": "Code review and debugging"
      }
    },
    {
      "id": "glm-5.3-flash-4bit",
      "name": "GLM 5.3 Flash",
      "provider": "MiaAI-Lab",
      "mark": "zai.svg",
      "description": "Start with the documented RoCE pair. MiaAI-Lab also documents a four-Spark ring.",
      "source": "https://github.com/MiaAI-Lab/GLM-5.3-Flash-DGX-Sparks-1104/blob/80be4d6882837d8e0840827ae2aab6124fd48da3/README.md#requirements",
      "test": null,
      "deployments": {
        "dgx-spark-pair": {
          "format": "EXL3 4-bit",
          "estimateGB": null,
          "weightGB": 175.6,
          "checkpoint": "https://huggingface.co/Mia-AiLab/GLM-5.3-Flash-EXL3-4bpw-TensorFold/tree/76c0b5173166d2795dd48860f45d8224817f894c",
          "runtime": "TensorFold · CUDA",
          "setup": "Start with the documented RoCE pair. MiaAI-Lab also documents a four-Spark ring.",
          "setupUrl": "https://github.com/MiaAI-Lab/GLM-5.3-Flash-DGX-Sparks-1104/blob/80be4d6882837d8e0840827ae2aab6124fd48da3/README.md#requirements",
          "source": "https://github.com/MiaAI-Lab/GLM-5.3-Flash-DGX-Sparks-1104/blob/80be4d6882837d8e0840827ae2aab6124fd48da3/README.md#requirements",
          "command": null,
          "creator": {
            "name": "Mia’s AI Lab",
            "website": "https://mia-ai.net/",
            "logo": "/assets/mia-ai-lab-logo.png",
            "article": "https://mia-ai.net/models/GLM-5.3-Flash-DGX-Sparks-1104",
            "podman": false
          },
          "engineVersion": "f350cd3e + Mia’s patches",
          "decode": {
            "rate": "47.8",
            "hardware": "Two DGX Sparks",
            "conditions": "One sampled prose reply with thinking on, EXL3 and DFlash2.",
            "source": "https://github.com/MiaAI-Lab/GLM-5.3-Flash-DGX-Sparks-1104/blob/80be4d6882837d8e0840827ae2aab6124fd48da3/README.md#performance",
            "credit": "Mia’s AI Lab"
          }
        },
        "dgx-spark-quad": {
          "format": "EXL3 4-bit",
          "estimateGB": null,
          "weightGB": 175.6,
          "checkpoint": "https://huggingface.co/Mia-AiLab/GLM-5.3-Flash-EXL3-4bpw-TensorFold/tree/76c0b5173166d2795dd48860f45d8224817f894c",
          "runtime": "TensorFold · CUDA",
          "setup": "Start with the documented RoCE pair. MiaAI-Lab also documents a four-Spark ring.",
          "setupUrl": "https://github.com/MiaAI-Lab/GLM-5.3-Flash-DGX-Sparks-1104/blob/80be4d6882837d8e0840827ae2aab6124fd48da3/README.md#requirements",
          "source": "https://github.com/MiaAI-Lab/GLM-5.3-Flash-DGX-Sparks-1104/blob/80be4d6882837d8e0840827ae2aab6124fd48da3/README.md#requirements",
          "command": null,
          "creator": {
            "name": "Mia’s AI Lab",
            "website": "https://mia-ai.net/",
            "logo": "/assets/mia-ai-lab-logo.png",
            "article": "https://mia-ai.net/models/GLM-5.3-Flash-DGX-Sparks-1104",
            "podman": false
          },
          "engineVersion": "f350cd3e + Mia’s patches",
          "decode": {
            "rate": "77.2",
            "hardware": "Four DGX Sparks",
            "conditions": "One sampled prose reply with thinking on, EXL3 and DFlash2 on a RoCE ring.",
            "source": "https://github.com/MiaAI-Lab/GLM-5.3-Flash-DGX-Sparks-1104/blob/80be4d6882837d8e0840827ae2aab6124fd48da3/README.md#performance-at-four-sparks",
            "credit": "Mia’s AI Lab"
          }
        }
      },
      "capabilities": {
        "use": "Start with repository tasks and tool-assisted analysis on a multi-GPU setup.",
        "source": "https://huggingface.co/zai-org/GLM-5.3-Flash",
        "epoch": {
          "name": "GLM-5.3-Flash",
          "score": 153.78,
          "low": 152.04,
          "high": 155.62
        },
        "size": "320B",
        "sizeLabel": "320 billion parameters",
        "sizeSource": "https://huggingface.co/zai-org/GLM-5.3-Flash",
        "shortUse": "Coding across a repository",
        "sizeDetails": "Uses 18B parameters at a time while generating text."
      }
    },
    {
      "id": "glm-5.3-flash-8bit",
      "name": "GLM 5.3 Flash",
      "provider": "MiaAI-Lab",
      "mark": "zai.svg",
      "description": "Uses Z.AI’s original FP8 weights. MiaAI-Lab documents this experimental setup on a four-Spark RoCE ring.",
      "source": "https://github.com/MiaAI-Lab/GLM-5.3-Flash-DGX-Sparks-1104/blob/80be4d6882837d8e0840827ae2aab6124fd48da3/README.md#block-fp8-weights-fp81-zais-own-release-four-sparks-only",
      "test": null,
      "deployments": {
        "dgx-spark-quad": {
          "format": "Block FP8",
          "estimateGB": null,
          "weightGB": 328.3,
          "checkpoint": "https://huggingface.co/zai-org/GLM-5.3-Flash/tree/46a050bec13002ad29ee795c16e61cf4fa3c4a9c",
          "runtime": "TensorFold · CUDA",
          "setup": "Uses Z.AI’s original FP8 weights. MiaAI-Lab documents this experimental setup on a four-Spark RoCE ring.",
          "setupUrl": "https://github.com/MiaAI-Lab/GLM-5.3-Flash-DGX-Sparks-1104/blob/80be4d6882837d8e0840827ae2aab6124fd48da3/README.md#block-fp8-weights-fp81-zais-own-release-four-sparks-only",
          "source": "https://github.com/MiaAI-Lab/GLM-5.3-Flash-DGX-Sparks-1104/blob/80be4d6882837d8e0840827ae2aab6124fd48da3/README.md#block-fp8-weights-fp81-zais-own-release-four-sparks-only",
          "command": null,
          "creator": {
            "name": "Mia’s AI Lab",
            "website": "https://mia-ai.net/",
            "logo": "/assets/mia-ai-lab-logo.png",
            "article": "https://mia-ai.net/models/GLM-5.3-Flash-DGX-Sparks-1104",
            "podman": false
          },
          "engineVersion": "f350cd3e + Mia’s patches",
          "decode": {
            "rate": "52.2",
            "hardware": "Four DGX Sparks",
            "conditions": "One sampled prose reply with thinking on, original block FP8 weights on a RoCE ring.",
            "source": "https://github.com/MiaAI-Lab/GLM-5.3-Flash-DGX-Sparks-1104/blob/80be4d6882837d8e0840827ae2aab6124fd48da3/README.md#performance-at-four-sparks",
            "credit": "Mia’s AI Lab"
          }
        }
      },
      "capabilities": {
        "use": "Start with repository tasks and tool-assisted analysis on a multi-GPU setup.",
        "source": "https://huggingface.co/zai-org/GLM-5.3-Flash",
        "epoch": {
          "name": "GLM-5.3-Flash",
          "score": 153.78,
          "low": 152.04,
          "high": 155.62
        },
        "size": "320B",
        "sizeLabel": "320 billion parameters",
        "sizeSource": "https://huggingface.co/zai-org/GLM-5.3-Flash",
        "shortUse": "Coding across a repository",
        "sizeDetails": "Uses 18B parameters at a time while generating text."
      }
    },
    {
      "id": "qwen3.8-27b-mlx-spark",
      "name": "Qwen3.8-27B",
      "provider": "Qwen",
      "mark": "qwen-color.svg",
      "description": "Mia’s original 4-bit configuration uses Docker, NVIDIA Container Toolkit, and a separate DFlash2 checkpoint.",
      "source": "https://github.com/MiaAI-Lab/Qwen3.8-27B-DGX-Spark-TensorFold/blob/c11ff138871cf43d12898d79dbfa4dc952d17ca4/README.md#quick-start",
      "test": null,
      "deployments": {
        "dgx-spark": {
          "format": "MLX 4-bit · DFlash2",
          "estimateGB": null,
          "weightGB": null,
          "checkpoint": "https://huggingface.co/Vontra/Qwen3.8-27B-MLX-4bit/tree/22d8d538154e0e5b6f8dcb1fcc74a60a0798104c",
          "runtime": "TensorFold · CUDA",
          "setup": "Mia’s original 4-bit configuration uses Docker, NVIDIA Container Toolkit, and a separate DFlash2 checkpoint.",
          "setupUrl": "https://github.com/MiaAI-Lab/Qwen3.8-27B-DGX-Spark-TensorFold/blob/c11ff138871cf43d12898d79dbfa4dc952d17ca4/README.md#quick-start",
          "source": "https://github.com/MiaAI-Lab/Qwen3.8-27B-DGX-Spark-TensorFold/blob/c11ff138871cf43d12898d79dbfa4dc952d17ca4/README.md#quick-start",
          "command": null,
          "creator": {
            "name": "Mia’s AI Lab",
            "website": "https://mia-ai.net/",
            "logo": "/assets/mia-ai-lab-logo.png",
            "article": "https://mia-ai.net/models/Qwen3.8-27B-DGX-Spark-TensorFold",
            "podman": false
          },
          "engineVersion": "0.6.0 + Mia’s patches",
          "decode": {
            "rate": "58.3",
            "hardware": "One DGX Spark",
            "conditions": "One prose reply, MLX 4-bit with DFlash2 and FP8 KV.",
            "source": "https://mia-ai.net/models/Qwen3.8-27B-DGX-Spark-TensorFold",
            "credit": "Mia’s AI Lab"
          }
        }
      },
      "codingReference": "https://huggingface.co/Qwen/Qwen3.8-27B#benchmark-results",
      "capabilities": {
        "use": "Start with code review, debugging and focused reasoning over documents.",
        "source": "https://huggingface.co/Qwen/Qwen3.8-27B#benchmark-results",
        "epoch": {
          "name": "Qwen 3.8 27B",
          "score": 152.65,
          "low": 150.7,
          "high": 154.76
        },
        "size": "27B",
        "sizeLabel": "27 billion parameters",
        "sizeSource": "https://huggingface.co/Qwen/Qwen3.8-27B#benchmark-results",
        "shortUse": "Code review and debugging"
      }
    },
    {
      "id": "ornith-1.5-122b",
      "name": "Ornith 1.5 122B",
      "provider": "MiaAI-Lab",
      "description": "Run Mia’s Ornith checkpoint and the incoai DFlash2 drafter on one Spark with her patched native TensorFold image.",
      "source": "https://github.com/MiaAI-Lab/Ornith-1.5-122B-TensorFold/blob/0c8fd6a0d1e2ece7557978a7d8c3d96d137eb075/README.md#quick-start",
      "test": null,
      "deployments": {
        "dgx-spark": {
          "format": "EXL3 4-bit · DFlash2",
          "estimateGB": null,
          "weightGB": null,
          "checkpoint": "https://huggingface.co/Mia-AiLab/Ornith-1.5-122B-EXL3-EAQv1-4-bpw",
          "runtime": "TensorFold · CUDA",
          "setup": "Run Mia’s Ornith checkpoint and the incoai DFlash2 drafter on one Spark with her patched native TensorFold image.",
          "setupUrl": "https://github.com/MiaAI-Lab/Ornith-1.5-122B-TensorFold/blob/0c8fd6a0d1e2ece7557978a7d8c3d96d137eb075/README.md#quick-start",
          "source": "https://github.com/MiaAI-Lab/Ornith-1.5-122B-TensorFold/blob/0c8fd6a0d1e2ece7557978a7d8c3d96d137eb075/README.md#quick-start",
          "command": null,
          "creator": {
            "name": "Mia’s AI Lab",
            "website": "https://mia-ai.net/",
            "logo": "/assets/mia-ai-lab-logo.png",
            "article": "https://mia-ai.net/",
            "podman": false
          },
          "engineVersion": "d1018b31 + Mia’s patches",
          "decode": {
            "rate": "52.6",
            "hardware": "One DGX Spark",
            "conditions": "One prose reply with EXL3 4-bit and DFlash2.",
            "source": "https://github.com/MiaAI-Lab/Ornith-1.5-122B-TensorFold/blob/0c8fd6a0d1e2ece7557978a7d8c3d96d137eb075/README.md#performance",
            "credit": "Mia’s AI Lab"
          }
        }
      },
      "capabilities": {
        "use": "Start with coding tasks you can grade with unit tests.",
        "source": "https://huggingface.co/Mia-AiLab/Ornith-1.5-122B-EXL3-EAQv1-4-bpw",
        "epoch": null,
        "size": "122B",
        "sizeLabel": "122 billion parameters",
        "sizeSource": "https://huggingface.co/ornith-ai/Ornith-1.5-122B",
        "shortUse": "Coding with unit tests"
      }
    },
    {
      "id": "nemotron-3.5-bolt",
      "name": "Nemotron 3.5 Bolt",
      "provider": "NVIDIA",
      "mark": "nvidia-color.svg",
      "description": "Use NVIDIA’s 74B-A9B checkpoint and LiLiCorr drafter on a single Spark. Mia’s guide covers Docker, CUDA and model storage.",
      "source": "https://github.com/MiaAI-Lab/Nemotron-3.5-Bolt-TensorFold-DGX-Spark/blob/6b064a31dd47e4f959e9041d6129eb2bad830481/README.md#quick-start",
      "test": null,
      "deployments": {
        "dgx-spark": {
          "format": "IQ2_XXS + NVFP4 · LiLiCorr",
          "estimateGB": null,
          "weightGB": null,
          "checkpoint": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB",
          "runtime": "TensorFold · CUDA",
          "setup": "Use NVIDIA’s 74B-A9B checkpoint and LiLiCorr drafter on a single Spark. Mia’s guide covers Docker, CUDA and model storage.",
          "setupUrl": "https://github.com/MiaAI-Lab/Nemotron-3.5-Bolt-TensorFold-DGX-Spark/blob/6b064a31dd47e4f959e9041d6129eb2bad830481/README.md#quick-start",
          "source": "https://github.com/MiaAI-Lab/Nemotron-3.5-Bolt-TensorFold-DGX-Spark/blob/6b064a31dd47e4f959e9041d6129eb2bad830481/README.md#quick-start",
          "command": null,
          "creator": {
            "name": "Mia’s AI Lab",
            "website": "https://mia-ai.net/",
            "logo": "/assets/mia-ai-lab-logo.png",
            "article": "https://mia-ai.net/models/Nemotron-3.5-Bolt-TensorFold-DGX-Spark",
            "podman": false
          },
          "engineVersion": "1.1.6 + Mia’s patches",
          "decode": {
            "rate": "73.8",
            "hardware": "One DGX Spark",
            "conditions": "One greedy prose reply, thinking off, with the LiLiCorr drafter.",
            "source": "https://github.com/MiaAI-Lab/Nemotron-3.5-Bolt-TensorFold-DGX-Spark/blob/6b064a31dd47e4f959e9041d6129eb2bad830481/README.md#speed",
            "credit": "Mia’s AI Lab"
          }
        }
      },
      "capabilities": {
        "use": "Start with tool-assisted coding and structured outputs. Check the recipe for input support.",
        "source": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB",
        "epoch": null,
        "size": "74B",
        "sizeLabel": "74 billion parameters",
        "sizeSource": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB",
        "shortUse": "Coding and structured output",
        "sizeDetails": "Uses 9B parameters at a time while generating text."
      }
    },
    {
      "id": "qwen3.8-flash-next-mia-dual",
      "name": "Qwen3.8 Flash Next",
      "provider": "Qwen",
      "mark": "qwen-color.svg",
      "description": "Use Mia’s setup for two Sparks linked over RoCE. Her guide covers both NVFP4 and INT4-AutoRound checkpoints and starts the two ranks.",
      "source": "https://github.com/MiaAI-Lab/Qwen3.8-Flash-Dual-DGX-Sparks-TensorFold/blob/dc05c912342ec19467b3b7b686888de763f2ada1/README.md#quick-start",
      "test": null,
      "deployments": {
        "dgx-spark-pair": {
          "format": "NVFP4 or INT4-AutoRound · MTP",
          "estimateGB": null,
          "weightGB": null,
          "checkpoint": "https://huggingface.co/nvidia/Qwen3.8-Flash-Next-NVFP4/tree/fc694b54fb0174e0913e6adf86691ef85a4ead47",
          "runtime": "TensorFold · CUDA",
          "setup": "Use Mia’s setup for two Sparks linked over RoCE. Her guide covers both NVFP4 and INT4-AutoRound checkpoints and starts the two ranks.",
          "setupUrl": "https://github.com/MiaAI-Lab/Qwen3.8-Flash-Dual-DGX-Sparks-TensorFold/blob/dc05c912342ec19467b3b7b686888de763f2ada1/README.md#quick-start",
          "source": "https://github.com/MiaAI-Lab/Qwen3.8-Flash-Dual-DGX-Sparks-TensorFold/blob/dc05c912342ec19467b3b7b686888de763f2ada1/README.md#quick-start",
          "command": null,
          "creator": {
            "name": "Mia’s AI Lab",
            "website": "https://mia-ai.net/",
            "logo": "/assets/mia-ai-lab-logo.png",
            "article": "https://mia-ai.net/models/Qwen3.8-Flash-Dual-DGX-Sparks-TensorFold",
            "podman": false
          },
          "engineVersion": "db281878 + Mia’s patches",
          "decode": {
            "rate": "67.4",
            "hardware": "Two DGX Sparks",
            "conditions": "One prose reply, NVIDIA NVFP4 with MTP, two ranks over RoCE.",
            "source": "https://mia-ai.net/models/Qwen3.8-Flash-Dual-DGX-Sparks-TensorFold",
            "credit": "Mia’s AI Lab"
          }
        }
      },
      "codingReference": "https://huggingface.co/Qwen/Qwen3.8-Flash-Next#benchmark-results",
      "capabilities": {
        "use": "Start with multi-file coding and long documents. Check the device recipe for image and video support.",
        "source": "https://huggingface.co/Qwen/Qwen3.8-Flash-Next#benchmark-results",
        "epoch": null,
        "size": "125B",
        "sizeLabel": "125 billion parameters in the main network",
        "sizeSource": "https://huggingface.co/Qwen/Qwen3.8-Flash-Next#benchmark-results",
        "shortUse": "Coding and long documents",
        "sizeDetails": "Qwen reports 6B active parameters, plus 51B for its text lookup table and 4B for predicting ahead.",
        "referenceCheck": "Qwen reports 62.5% on its corrected SWE-bench Pro task set using Claude Code at 256K context."
      }
    },
    {
      "id": "deepseek-v4.1-flash-1017-exl3",
      "name": "DeepSeek V4.1 Flash 1017",
      "provider": "DeepSeek",
      "description": "Use Mia’s EXL3 checkpoint on two Sparks with a ConnectX-7 link. Her guide covers the TensorFold image, weights and launch on both nodes.",
      "source": "https://github.com/MiaAI-Lab/DeepSeek-v4.1-Flash-1017-2x-DGX-Sparks/blob/c7f2b85100f0a7b099c92a4e21825e30d93c9c51/README.md#quick-start",
      "test": null,
      "deployments": {
        "dgx-spark-pair": {
          "format": "EXL3 EAQ 2.9 bpw · DSpark",
          "estimateGB": null,
          "weightGB": null,
          "checkpoint": "https://huggingface.co/Mia-AiLab/DeepSeek-V4.1-Flash-1017-EXL3-EAQ-v1-2.9-bpw",
          "runtime": "TensorFold · CUDA",
          "setup": "Use Mia’s EXL3 checkpoint on two Sparks with a ConnectX-7 link. Her guide covers the TensorFold image, weights and launch on both nodes.",
          "setupUrl": "https://github.com/MiaAI-Lab/DeepSeek-v4.1-Flash-1017-2x-DGX-Sparks/blob/c7f2b85100f0a7b099c92a4e21825e30d93c9c51/README.md#quick-start",
          "source": "https://github.com/MiaAI-Lab/DeepSeek-v4.1-Flash-1017-2x-DGX-Sparks/blob/c7f2b85100f0a7b099c92a4e21825e30d93c9c51/README.md#quick-start",
          "command": null,
          "creator": {
            "name": "Mia’s AI Lab",
            "website": "https://mia-ai.net/",
            "logo": "/assets/mia-ai-lab-logo.png",
            "article": "https://mia-ai.net/",
            "podman": false
          },
          "engineVersion": "fbbc9713 + recipe patches",
          "decode": {
            "rate": "63.9",
            "hardware": "Two DGX Sparks",
            "conditions": "One sampled prose reply with thinking on, EXL3 and DSpark.",
            "source": "https://github.com/MiaAI-Lab/DeepSeek-v4.1-Flash-1017-2x-DGX-Sparks/blob/c7f2b85100f0a7b099c92a4e21825e30d93c9c51/README.md#performance",
            "credit": "Mia’s AI Lab"
          }
        }
      },
      "codingReference": "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017",
      "capabilities": {
        "use": "Start with multi-step coding, repository analysis and tool-driven research.",
        "source": "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017",
        "epoch": {
          "name": "DeepSeek V4.1 Flash 1017",
          "score": 159.4,
          "low": 157.16,
          "high": 161.93
        },
        "size": "552B",
        "sizeLabel": "552 billion parameters in the main network",
        "sizeSource": "https://arxiv.org/html/2609.19969v1",
        "shortUse": "Multi-step coding and analysis",
        "sizeDetails": "The main network has 552B parameters, plus 196B in a lookup table. It uses 16B parameters at a time while generating text."
      }
    },
    {
      "id": "glm-5.3-exl3",
      "name": "GLM 5.3",
      "provider": "Z.AI",
      "mark": "zai.svg",
      "description": "Run the full GLM 5.3 on three Sparks. Mia’s guide covers the three-node network and sharing the checkpoint over NFS.",
      "source": "https://github.com/MiaAI-Lab/GLM-5.3-EXL3-3x-DGX-Sparks-TensorFold/blob/eb6414b488c3bc58605cad3dc96391c7584d6bec/README.md#quick-start",
      "test": null,
      "deployments": {
        "dgx-spark-trio": {
          "format": "EXL3 2.75 bpw · DSpark",
          "estimateGB": null,
          "weightGB": null,
          "checkpoint": "https://huggingface.co/Mia-AiLab/GLM-5.3-EXL3-2.75bpw-TensorFold",
          "runtime": "TensorFold · CUDA",
          "setup": "Run the full GLM 5.3 on three Sparks. Mia’s guide covers the three-node network and sharing the checkpoint over NFS.",
          "setupUrl": "https://github.com/MiaAI-Lab/GLM-5.3-EXL3-3x-DGX-Sparks-TensorFold/blob/eb6414b488c3bc58605cad3dc96391c7584d6bec/README.md#quick-start",
          "source": "https://github.com/MiaAI-Lab/GLM-5.3-EXL3-3x-DGX-Sparks-TensorFold/blob/eb6414b488c3bc58605cad3dc96391c7584d6bec/README.md#quick-start",
          "command": null,
          "creator": {
            "name": "Mia’s AI Lab",
            "website": "https://mia-ai.net/",
            "logo": "/assets/mia-ai-lab-logo.png",
            "article": "https://mia-ai.net/models/GLM-5.3-EXL3-3x-DGX-Sparks-TensorFold",
            "podman": false
          },
          "engineVersion": "0.6.0 + Mia’s patches",
          "decode": {
            "rate": "31.3",
            "hardware": "Three DGX Sparks",
            "conditions": "One prose reply, full GLM 5.3 at EXL3 2.75 bpw with DSpark.",
            "source": "https://mia-ai.net/models/GLM-5.3-EXL3-3x-DGX-Sparks-TensorFold",
            "credit": "Mia’s AI Lab"
          }
        }
      },
      "capabilities": {
        "use": "Start with larger repository tasks and multi-step coding workflows.",
        "source": "https://huggingface.co/zai-org/GLM-5.3",
        "epoch": {
          "name": "GLM-5.3",
          "score": 157.1,
          "low": 155.04,
          "high": 159.33
        },
        "size": "744B",
        "sizeLabel": "744 billion parameters",
        "sizeSource": "https://huggingface.co/zai-org/GLM-5.3",
        "shortUse": "Large coding projects"
      }
    },
    {
      "id": "deepseek-v4.1-flash-1017-fp8",
      "name": "DeepSeek V4.1 Flash 1017",
      "provider": "DeepSeek",
      "description": "Serve DeepSeek’s original checkpoint across a four-Spark RoCE ring using Mia’s TensorFold recipe.",
      "source": "https://github.com/MiaAI-Lab/DeepSeek-V4.1-Flash-1017-4x-DGX-Sparks-TensorFold/blob/8ffa67d53ef15bab997399135f85e8b0910f30be/README.md#quick-start",
      "test": null,
      "deployments": {
        "dgx-spark-quad": {
          "format": "FP8 + MXFP4 · DSpark",
          "estimateGB": null,
          "weightGB": null,
          "checkpoint": "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017",
          "runtime": "TensorFold · CUDA",
          "setup": "Serve DeepSeek’s original checkpoint across a four-Spark RoCE ring using Mia’s TensorFold recipe.",
          "setupUrl": "https://github.com/MiaAI-Lab/DeepSeek-V4.1-Flash-1017-4x-DGX-Sparks-TensorFold/blob/8ffa67d53ef15bab997399135f85e8b0910f30be/README.md#quick-start",
          "source": "https://github.com/MiaAI-Lab/DeepSeek-V4.1-Flash-1017-4x-DGX-Sparks-TensorFold/blob/8ffa67d53ef15bab997399135f85e8b0910f30be/README.md#quick-start",
          "command": null,
          "creator": {
            "name": "Mia’s AI Lab",
            "website": "https://mia-ai.net/",
            "logo": "/assets/mia-ai-lab-logo.png",
            "article": "https://mia-ai.net/models/DeepSeek-V4.1-Flash-1017-4x-DGX-Sparks-TensorFold",
            "podman": false
          },
          "engineVersion": "59402086 + recipe patches",
          "decode": {
            "rate": "65.8",
            "hardware": "Four DGX Sparks",
            "conditions": "One prose reply with the original DeepSeek checkpoint and DSpark.",
            "source": "https://mia-ai.net/models/DeepSeek-V4.1-Flash-1017-4x-DGX-Sparks-TensorFold",
            "credit": "Mia’s AI Lab"
          }
        }
      },
      "codingReference": "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017",
      "capabilities": {
        "use": "Start with multi-step coding, repository analysis and tool-driven research.",
        "source": "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017",
        "epoch": {
          "name": "DeepSeek V4.1 Flash 1017",
          "score": 159.4,
          "low": 157.16,
          "high": 161.93
        },
        "size": "552B",
        "sizeLabel": "552 billion parameters in the main network",
        "sizeSource": "https://arxiv.org/html/2609.19969v1",
        "shortUse": "Multi-step coding and analysis",
        "sizeDetails": "The main network has 552B parameters, plus 196B in a lookup table. It uses 16B parameters at a time while generating text."
      }
    },
    {
      "id": "qwen3.8-flash-next-rtx-pro-6000",
      "name": "Qwen3.8 Flash Next",
      "provider": "Qwen",
      "mark": "qwen-color.svg",
      "description": "For one 96 GB RTX PRO 6000 Blackwell on Linux x86_64. Use Mia’s Docker build and keep the checkpoint on an NVMe drive.",
      "source": "https://github.com/MiaAI-Lab/Qwen3.8-Flash-Next-RTX-PRO-6000-TernsorFold/blob/57d815c2c2415fcc4f3cc5470e23746199c0abc1/README.md#quick-start",
      "test": null,
      "deployments": {
        "rtx-pro-6000": {
          "format": "NVIDIA FP4 QAD · MTP",
          "estimateGB": null,
          "weightGB": null,
          "checkpoint": "https://huggingface.co/nvidia/Qwen3.8-Flash-Next-FP4-Local/tree/b9ce18a6887e18c089fdea8509a33e63429a3300",
          "runtime": "TensorFold · CUDA",
          "setup": "For one 96 GB RTX PRO 6000 Blackwell on Linux x86_64. Use Mia’s Docker build and keep the checkpoint on an NVMe drive.",
          "setupUrl": "https://github.com/MiaAI-Lab/Qwen3.8-Flash-Next-RTX-PRO-6000-TernsorFold/blob/57d815c2c2415fcc4f3cc5470e23746199c0abc1/README.md#quick-start",
          "source": "https://github.com/MiaAI-Lab/Qwen3.8-Flash-Next-RTX-PRO-6000-TernsorFold/blob/57d815c2c2415fcc4f3cc5470e23746199c0abc1/README.md#quick-start",
          "command": null,
          "creator": {
            "name": "Mia’s AI Lab",
            "website": "https://mia-ai.net/",
            "logo": "/assets/mia-ai-lab-logo.png",
            "article": "https://mia-ai.net/models/Qwen3.8-Flash-Next-RTX-PRO-6000-TernsorFold",
            "podman": false
          },
          "engineVersion": "1.1.4 + Mia’s patches",
          "decode": {
            "rate": "271.4",
            "hardware": "One RTX PRO 6000 Blackwell",
            "conditions": "One prose reply at 400 W, NVIDIA FP4 QAD and FP8 KV.",
            "source": "https://mia-ai.net/models/Qwen3.8-Flash-Next-RTX-PRO-6000-TernsorFold",
            "credit": "Mia’s AI Lab"
          }
        }
      },
      "codingReference": "https://huggingface.co/Qwen/Qwen3.8-Flash-Next#benchmark-results",
      "capabilities": {
        "use": "Start with multi-file coding and long documents. Check the device recipe for image and video support.",
        "source": "https://huggingface.co/Qwen/Qwen3.8-Flash-Next#benchmark-results",
        "epoch": null,
        "size": "125B",
        "sizeLabel": "125 billion parameters in the main network",
        "sizeSource": "https://huggingface.co/Qwen/Qwen3.8-Flash-Next#benchmark-results",
        "shortUse": "Coding and long documents",
        "sizeDetails": "Qwen reports 6B active parameters, plus 51B for its text lookup table and 4B for predicting ahead.",
        "referenceCheck": "Qwen reports 62.5% on its corrected SWE-bench Pro task set using Claude Code at 256K context."
      }
    },
    {
      "id": "glm-5.3-flash-rtx-pro-6000-pair",
      "name": "GLM 5.3 Flash",
      "provider": "Z.AI",
      "mark": "zai.svg",
      "description": "Mia’s current native TensorFold setup uses two 96 GB RTX PRO 6000 Blackwell cards in one Linux host.",
      "source": "https://github.com/MiaAI-Lab/GLM-5.3-Flash-2x-RTX-PRO-6000-TensorFold/blob/fc1ee3a5381031e2357662b46c2fb44b20190429/README.md#quick-start",
      "test": null,
      "deployments": {
        "rtx-pro-6000-pair": {
          "format": "EXL3 4-bit",
          "estimateGB": null,
          "weightGB": null,
          "checkpoint": "https://huggingface.co/Mia-AiLab/GLM-5.3-Flash-EXL3-4bpw-TensorFold/tree/76c0b5173166d2795dd48860f45d8224817f894c",
          "runtime": "TensorFold · CUDA",
          "setup": "Mia’s current native TensorFold setup uses two 96 GB RTX PRO 6000 Blackwell cards in one Linux host.",
          "setupUrl": "https://github.com/MiaAI-Lab/GLM-5.3-Flash-2x-RTX-PRO-6000-TensorFold/blob/fc1ee3a5381031e2357662b46c2fb44b20190429/README.md#quick-start",
          "source": "https://github.com/MiaAI-Lab/GLM-5.3-Flash-2x-RTX-PRO-6000-TensorFold/blob/fc1ee3a5381031e2357662b46c2fb44b20190429/README.md#quick-start",
          "command": null,
          "creator": {
            "name": "Mia’s AI Lab",
            "website": "https://mia-ai.net/",
            "logo": "/assets/mia-ai-lab-logo.png",
            "article": "https://mia-ai.net/models/GLM-5.3-Flash-2x-RTX-PRO-6000-TensorFold",
            "podman": false
          },
          "engineVersion": "f350cd3e + Mia’s patches",
          "decode": {
            "rate": "245.5",
            "hardware": "Two RTX PRO 6000 Blackwell GPUs",
            "conditions": "One sampled prose reply at 400 W per GPU, EXL3 with DFlash2.",
            "source": "https://mia-ai.net/models/GLM-5.3-Flash-2x-RTX-PRO-6000-TensorFold",
            "credit": "Mia’s AI Lab"
          }
        }
      },
      "capabilities": {
        "use": "Start with repository tasks and tool-assisted analysis on a multi-GPU setup.",
        "source": "https://huggingface.co/zai-org/GLM-5.3-Flash",
        "epoch": {
          "name": "GLM-5.3-Flash",
          "score": 153.78,
          "low": 152.04,
          "high": 155.62
        },
        "size": "320B",
        "sizeLabel": "320 billion parameters",
        "sizeSource": "https://huggingface.co/zai-org/GLM-5.3-Flash",
        "shortUse": "Coding across a repository",
        "sizeDetails": "Uses 18B parameters at a time while generating text."
      }
    },
    {
      "id": "glm-5.3-flash-rtx-pro-6000-pair-aevonix",
      "name": "GLM 5.3 Flash",
      "provider": "Z.AI",
      "mark": "zai.svg",
      "description": "Mia and Aevonix Research documented this two-GPU setup. The write-up is public and its original recipe repository is currently unavailable.",
      "source": "https://mia-ai.net/models/GLM-5.3-Flash-EXL3-2x-RTX-PRO-6000-TensorFold",
      "test": null,
      "deployments": {
        "rtx-pro-6000-pair": {
          "format": "EXL3 4-bit · Aevonix setup",
          "estimateGB": null,
          "weightGB": null,
          "checkpoint": null,
          "runtime": "TensorFold · CUDA",
          "setup": "Mia and Aevonix Research documented this two-GPU setup. The write-up is public and its original recipe repository is currently unavailable.",
          "setupUrl": "https://mia-ai.net/models/GLM-5.3-Flash-EXL3-2x-RTX-PRO-6000-TensorFold",
          "source": "https://mia-ai.net/models/GLM-5.3-Flash-EXL3-2x-RTX-PRO-6000-TensorFold",
          "command": null,
          "creator": {
            "name": "Mia’s AI Lab",
            "website": "https://mia-ai.net/",
            "logo": "/assets/mia-ai-lab-logo.png",
            "article": "https://mia-ai.net/models/GLM-5.3-Flash-EXL3-2x-RTX-PRO-6000-TensorFold",
            "podman": false
          },
          "engineVersion": null,
          "guideLabel": "Read Mia’s write-up ↗"
        }
      },
      "capabilities": {
        "use": "Start with repository tasks and tool-assisted analysis on a multi-GPU setup.",
        "source": "https://huggingface.co/zai-org/GLM-5.3-Flash",
        "epoch": {
          "name": "GLM-5.3-Flash",
          "score": 153.78,
          "low": 152.04,
          "high": 155.62
        },
        "size": "320B",
        "sizeLabel": "320 billion parameters",
        "sizeSource": "https://huggingface.co/zai-org/GLM-5.3-Flash",
        "shortUse": "Coding across a repository",
        "sizeDetails": "Uses 18B parameters at a time while generating text."
      }
    },
    {
      "id": "glm-5.3-flash-rtx-pro-6000-quad",
      "name": "GLM 5.3 Flash",
      "provider": "Z.AI",
      "mark": "zai.svg",
      "description": "Mia and Aevonix Research documented this four-GPU setup. The write-up is public and its original recipe repository is currently unavailable.",
      "source": "https://mia-ai.net/models/GLM-5.3-Flash-EXL3-4x-RTX-PRO-6000-TensorFold",
      "test": null,
      "deployments": {
        "rtx-pro-6000-quad": {
          "format": "EXL3 4-bit · Aevonix setup",
          "estimateGB": null,
          "weightGB": null,
          "checkpoint": null,
          "runtime": "TensorFold · CUDA",
          "setup": "Mia and Aevonix Research documented this four-GPU setup. The write-up is public and its original recipe repository is currently unavailable.",
          "setupUrl": "https://mia-ai.net/models/GLM-5.3-Flash-EXL3-4x-RTX-PRO-6000-TensorFold",
          "source": "https://mia-ai.net/models/GLM-5.3-Flash-EXL3-4x-RTX-PRO-6000-TensorFold",
          "command": null,
          "creator": {
            "name": "Mia’s AI Lab",
            "website": "https://mia-ai.net/",
            "logo": "/assets/mia-ai-lab-logo.png",
            "article": "https://mia-ai.net/models/GLM-5.3-Flash-EXL3-4x-RTX-PRO-6000-TensorFold",
            "podman": false
          },
          "engineVersion": null,
          "guideLabel": "Read Mia’s write-up ↗"
        }
      },
      "capabilities": {
        "use": "Start with repository tasks and tool-assisted analysis on a multi-GPU setup.",
        "source": "https://huggingface.co/zai-org/GLM-5.3-Flash",
        "epoch": {
          "name": "GLM-5.3-Flash",
          "score": 153.78,
          "low": 152.04,
          "high": 155.62
        },
        "size": "320B",
        "sizeLabel": "320 billion parameters",
        "sizeSource": "https://huggingface.co/zai-org/GLM-5.3-Flash",
        "shortUse": "Coding across a repository",
        "sizeDetails": "Uses 18B parameters at a time while generating text."
      }
    }
  ],
  "comparison": {
    "author": "Epoch AI",
    "citation": "Epoch Capabilities Index",
    "source": "https://epoch.ai/eci",
    "license": "https://creativecommons.org/licenses/by/4.0/",
    "retrieved": "2026-11-18",
    "data": "https://epoch.ai/data/eci_scores.csv",
    "sha256": "a8282e62c93c25d95a3921d9a8d4b90bcf05982634d4362a9e056244f1563309",
    "hosted": [
      {
        "name": "Claude Opus 5.5",
        "score": 168.7,
        "low": 165.87,
        "high": 171.89
      },
      {
        "name": "GPT-6.2 Astra",
        "score": 174.17,
        "low": 170.29,
        "high": 178.58
      },
      {
        "name": "Gemini 4 Carbon",
        "score": 169.78,
        "low": 166.59,
        "high": 173.31
      },
      {
        "name": "Claude Fable 5.2",
        "score": 173.5,
        "low": 170.09,
        "high": 177.36
      }
    ]
  },
  "multiDevice": [
    {
      "id": "glm-5.3-flash-4bit",
      "name": "GLM 5.3 Flash",
      "provider": "MiaAI-Lab",
      "mark": "zai.svg",
      "format": "EXL3 4-bit",
      "hardware": "2 or 4 DGX Sparks",
      "targets": [
        "dgx-spark-pair",
        "dgx-spark-quad"
      ],
      "article": "/models/GLM-5.3-Flash-DGX-Sparks-1104",
      "weightGB": 175.6,
      "description": "Start with the documented RoCE pair. MiaAI-Lab also documents a four-Spark ring.",
      "source": "https://github.com/MiaAI-Lab/GLM-5.3-Flash-DGX-Sparks-1104/blob/80be4d6882837d8e0840827ae2aab6124fd48da3/README.md#requirements",
      "checkpoint": "https://huggingface.co/Mia-AiLab/GLM-5.3-Flash-EXL3-4bpw-TensorFold/tree/76c0b5173166d2795dd48860f45d8224817f894c"
    },
    {
      "id": "glm-5.3-flash-8bit",
      "name": "GLM 5.3 Flash",
      "provider": "MiaAI-Lab",
      "mark": "zai.svg",
      "format": "Block FP8",
      "hardware": "4 DGX Sparks",
      "targets": [
        "dgx-spark-quad"
      ],
      "article": "/models/GLM-5.3-Flash-DGX-Sparks-1104",
      "weightGB": 328.3,
      "description": "Uses Z.AI’s original FP8 weights. MiaAI-Lab documents this experimental setup on a four-Spark RoCE ring.",
      "source": "https://github.com/MiaAI-Lab/GLM-5.3-Flash-DGX-Sparks-1104/blob/80be4d6882837d8e0840827ae2aab6124fd48da3/README.md#block-fp8-weights-fp81-zais-own-release-four-sparks-only",
      "checkpoint": "https://huggingface.co/zai-org/GLM-5.3-Flash/tree/46a050bec13002ad29ee795c16e61cf4fa3c4a9c"
    }
  ],
  "initialModel": "qwen3.8-27b"
}
