GLM 5.3 Flash · EXL3 4-bit
2 or 4 DGX Sparks · 175.6 GB weights
Start with the documented RoCE pair. MiaAI-Lab also documents a four-Spark ring.
Recipe by Mia’s AI Lab ↗
Podman launcher by TensorFold
Choose your computer, pick a model, and copy the command to run it.
Memory estimates use the default launch settings.
Start with code review, debugging and focused reasoning over documents.
Open setup guide ↗
Recipe by Mia’s AI Lab ↗
Podman launcher by TensorFold
(
set -e
export HF_HUB_CACHE="${HF_HUB_CACHE:-${HF_HOME:-$HOME/.cache/huggingface}/hub}"
tensorfold pull TensorFold/Qwen3.8-27B-MLX-4bit@22d8d538154e0e5b6f8dcb1fcc74a60a0798104c
tensorfold pull z-lab/Qwen3.8-27B-DFlash2@50307d4c4cde6860d4eee73e2547cd786fe8e8a4
tensorfold serve \
"$HF_HUB_CACHE/models--TensorFold--Qwen3.8-27B-MLX-4bit/snapshots/22d8d538154e0e5b6f8dcb1fcc74a60a0798104c" \
--drafter "$HF_HUB_CACHE/models--z-lab--Qwen3.8-27B-DFlash2/snapshots/50307d4c4cde6860d4eee73e2547cd786fe8e8a4" \
--name qwen3.8-27b --host 127.0.0.1 --port 8080 \
--context 8192 --parallel 1
)
Custom settings can change memory use and speed. The memory estimate and benchmark use their published settings.
Checkpoint task test Test details ↗
A task-quality measurement for this setup is pending.
This score combines several tests. A higher score means better results across those tests. The scores here describe the reference models. The local download and hosted app can give different results.
| Reference model | ECI | Credible interval |
|---|---|---|
| Qwen 3.8 27B | 152.7 | 150.7–154.8 |
| Claude Opus 5.5 | 168.7 | 165.9–171.9 |
| GPT-6.2 Astra | 174.2 | 170.3–178.6 |
| Gemini 4 Carbon | 169.8 | 166.6–173.3 |
| Claude Fable 5.2 | 173.5 | 170.1–177.4 |
No matching ECI score for this model in this snapshot.
Selected from Epoch Capabilities Index by Epoch AI ↗, retrieved 18 Nov 2026. CC BY 4.0 ↗
Model evidence ↗Engine target · TensorFold 1.2.9
TensorFold CLI on an Apple Silicon Mac. The engine recipe documents M5 Max and M3 Ultra runs.
Decode measurement pending for this setup.
Update TensorFold through its installation channel. The recipe uses the existing Hugging Face cache for matching weights.
2 or 4 DGX Sparks · 175.6 GB weights
Start with the documented RoCE pair. MiaAI-Lab also documents a four-Spark ring.
Recipe by Mia’s AI Lab ↗
Podman launcher by TensorFold
4 DGX Sparks · 328.3 GB weights
Uses Z.AI’s original FP8 weights. MiaAI-Lab documents this experimental setup on a four-Spark RoCE ring.
Recipe by Mia’s AI Lab ↗
Podman launcher by TensorFold