Apple Silicon ML, without Python.

Pre-1.0. Source private today, available for review on request. The core packages are the intended v1.0.0 compatibility surface. Used in production by skiff for local inference.

• Core MLX runtime — arrays, lazy evaluation, autograd, and compiled graphs, exposed as small Go packages (mlx, mlx/nn, mlx/optimizer, mlx/compile, mlx/fast).
• No C toolchain — binds the MLX libraries through purego FFI via apple. Single static binary, no cgo in the build.
• Model I/O — streaming safetensors reading and writing, weight discovery, and quantized-weight handling.
• Conventions kept — layers, optimizers, and fused fast-path kernels track the upstream Python and Swift MLX conventions, so ports stay legible.
• Profiling hooks — GPU trace capture that pairs with gputrace for source-level Metal attribution.

mlx-go provides the array runtime and stops there. Language models live in mlx-go-lm, distributed-training math in mlx-go-ccl, and transport in mlx-go-iroh. skiff uses the runtime for local inference without deploying Python; mlx-go-iroh is the only one of these modules published and tagged on GitHub today.

• mlx-go-lm — language models: inference, chat, serving, and training. Pre-1.0.
• mlx-go-iroh — the mesh transport substrate. Open source, tagged v0.3.0.
• mlx-go-ccl — collective communication and the DiLoCo outer loop. Experimental.
• mlx-go-ane — Apple Neural Engine execution paths. Open-source experiment.
• mlx-go-fedistill — delta-only federated distillation rounds. Research.
• mlx-go-tallytrain — federated distillation by one-byte votes. Research.
• mlx-go-vibethinker — the VibeThinker post-training recipe, reproduced. Open source.
• mlx-go-sia — the SIA self-improving-agent loop, ported. Open source.