ashhart/TensorFold

TensorFold is an inference engine written in Zig that directly reads model weights, performs tokenization, and invokes Metal or CUDA kernels. It can be installed via Homebrew on Mac and via release packages on Linux, providing a local service interface. The model table in the README specifies which hardware each model has been verified on, for example, Qwen3.5-2B on M5 Max and GB10, and Qwen3.8-27B on M5 Max and M3 Ultra. It also features an experimental capability called Living Weights, which writes new facts into the model's own weights during runtime and allows for one-click rollback.

TensorFold is an inference engine written in Zig that directly reads model weights, performs tokenization, and invokes Metal or CUDA kernels. It can be installed via Homebrew on Mac and via release packages on Linux, providing a local service interface. The model table in the README specifies which hardware each model has been verified on, for example, Qwen3.5-2B on M5 Max and GB10, and Qwen3.8-27B on M5 Max and M3 Ultra. It also features an experimental capability called Living Weights, which writes new facts into the model's own weights during runtime and allows for one-click rollback.

Why it matters A single binary directly reads weights and runs Metal or CUDA kernels, eliminating the need to configure Python and MLX environments during deployment.

Stars 1942 (+54 / 24h)

Reasons

  • 54 new stars in 24 hours (indicating attention only)
  • The README distinguishes between hardware that has been verified and hardware where only code exists without execution records; untested hardware is not listed as supported

Usage

  • Mac: brew install ashhart/tensorfold/tensorfold, then use tensorfold pull to download the pinned versions of models listed in the README, and tensorfold serve to start the service

Risks

  • Linux requires NVIDIA R580 or higher drivers (CUDA 13); older drivers will be rejected at startup
  • Living Weights is an experimental feature; the author notes that related facts may interfere with each other, and recall degrades after learning dozens of facts

Open

Trendshift(GitHub Trending 快照)

Explore the latest AI Radar updates