TensorFold

TensorFold is a native Zig inference engine that talks directly to Metal or CUDA. It reads supported MLX-format checkpoints without requiring a Python or MLX runtime for serving. Its local HTTP service exposes OpenAI-compatible and Anthropic-compatible routes, with compatibility limited by the documented model and platform matrix.

The Apache-2.0 engine’s 1.0.10 release adds Mac support for specified mixed-width checkpoints. Model weights retain their own licenses and hardware requirements. Homebrew installation and prebuilt releases are documented; the service defaults to loopback access, while remote exposure requires deliberate network and authentication configuration. Performance figures in the project documentation are maintainer or contributor measurements, not AITrending benchmarks.