874tokens / second
Nemotron 3.5 Lightning on RTX PRO 6000 Blackwell
Greedy code decode with drafting enabled, an MLX affine 4-bit checkpoint and the native CUDA engine.
v1.0.6 release measurementCompany & technical brief
Native inference.
On your infrastructure.
TensorFold runs language models on Apple Silicon and qualified NVIDIA GPUs, with a native Zig engine, parallel decoding and reusable prompt state.
01
The engine
Direct Metal and CUDA execution, written in Zig. TensorFold serves supported checkpoints through APIs compatible with OpenAI and Anthropic, on hardware teams own and operate.
Its supported paths include speculative decoding, prompt-state reuse and concurrent serving. Selected models can run across two Macs with MCDMA. Model support, hardware requirements and limitations are published for specific tested combinations.
02
Measured results
90 to 110milliseconds to resume
Nemotron conversation resume on GB10
Turns two to five with a kept prompt state on native CUDA. The same release notes report 0.15 to 0.30 seconds for version 1.0.7.
v1.0.8 release measurement199tokens / second, aggregate
GLM-5.3-Flash on two M5 Ultras
Eight concurrent requests sharing two Macs. The release records each stream matching its solo run.
v1.0.2 release measurementThese are separate release measurements on the stated workloads and hardware; results depend on the model, settings and machine.
03Project
Project
collaborations
- NVIDIA Inception
- TensorFold is a member of NVIDIA Inception, NVIDIA's program for startups.
- Mia's AI Lab
- We partnered with Mia and her team to optimise TensorFold recipes, with their patches and testing contributing to the published results.
- Satellite deployment
- TensorFold is being prepared for a satellite deployment. Hardware results will follow once the deployment has been qualified.
04
Start a conversation
Investment & partnerships
Contact us to discuss TensorFold's development and plans for hardware testing.
Contact TensorFold