Open source · 2026–present

Zynfer — from-scratch LLM inference in Zig

Narrow inference runtime aimed at AMD RDNA 4 (gfx1201) and Qwen; Apple Metal path is complete through Qwen3-4B int8.

  • Wrote CPU f32 reference ops, Metal kernels, KV cache, tokenizer/sampling, and a quantized Qwen3-4B Metal schedule.
  • Next work is the AMD gfx1201 / HIP path — not a universal ML framework.
AI Zig GPU Metal HIP OSX Linux Inference

github