Local inference. Open source.

Run LLMs, Fast.

Run open models on the GPUs you own, with native Metal and CUDA inference.

Beta 10 · Engine 1.1.4 · Apple silicon · macOS 13.3+

Install the CLIMetal · CUDA

curl -fsSL https://tensorfold.dev/install.sh | sh

Apache-2.0OpenAI-compatibleAnthropic API

Apple Silicon

NVIDIA CUDA

Models running with TensorFold

Inference,
measured.

Real runs on named hardware.
Open the release. Inspect the conditions.

Recorded performanceRelease 1.0.6

874tokens per second

RTX PRO 6000 Blackwell

Greedy code decode with drafting enabled, an MLX affine 4-bit checkpoint and the CUDA engine.

Each number is its own release measurement, on the model and machine named under it: not a like-for-like comparison.

Drafted tokens, retained prompts and connected machines, working through the same engine.

On supported models, a drafter proposes several tokens and the target checks them together, committing only the matching prefix.

Explore real community runs on Apple Silicon and NVIDIA GPUs, inspect the receipts, and share what your machine can do.

Inception member

Part of NVIDIA's startup program.

Recipes developed with Mia's AI Lab

With thanks to Mia and her team for their patches, setup and testing.

Deployment in preparation

TensorFold is being prepared for a satellite deployment, where every hardware resource counts.

Concept illustration. Hardware results will follow qualification.