Open Brain
GitHubGitHub RepoNeeds Review

ashhart/TensorFold

TensorFold is a Zig-based local inference server for Apple Silicon and NVIDIA GPUs; its contributor notes also ask coding agents to coordinate work before editing.

GitHub
Open Source Open MarkdownOpen JSON
Source Summary

TensorFold serves language models from a **Zig binary** using Metal or CUDA kernels, with no Python or MLX installation required for serving. Its README also lists Vulkan support and OpenAI-compatible, Responses, and Anthropic Messages APIs. For a coding-agent workflow, the project asks contributors to check Agent Traffic Control before editing, claim an available issue, and follow its contributor instructions. A pinned Qwen3.5-2B checkpoint is provided as a local-serving starting point. Hardware and model support vary by checkpoint and backend, and some paths are explicitly unqualified or experimental. The README advises checking the checkpoint’s memory floor and the documented hardware scope before serving.

Practical Implication

No separate implication has been recorded yet.

Agent-Ready Context
TensorFold serves language models from a **Zig binary** using Metal or CUDA kernels, with no Python or MLX installation required for serving. Its README also lists Vulkan support and OpenAI-compatible, Responses, and Anthropic Messages APIs.
For a coding-agent workflow, the project asks contributors to check Agent Traffic Control before editing, claim an available issue, and follow its contributor instructions. A pinned Qwen3.5-2B checkpoint is provided as a local-serving starting point.
Hardware and model support vary by checkpoint and backend, and some paths are explicitly unqualified or experimental. The README advises checking the checkpoint’s memory floor and the documented hardware scope before serving.
Connected Context · Feed7 Judgment

TensorFold adds a Zig-based serving option for local models, with Metal and CUDA kernels and compatible APIs. Compared with the supplied candidates, it offers a focused server rather than a bundled agent stack or Apple-specific agent-serving features. Its hardware and checkpoint caveats leave performance and deployment fit to be checked for each setup.

Context Map
infracoding#open-models#sandboxing