The repository is public, ungated, tagged for MLX and Apple Silicon, and lists a total size of 15.7 GiB.
The mixed layout uses 3-bit weights for MLP gate and up projections, 4-bit weights for attention, down_proj, and embeddings, and 6-bit weights for lm_head. The author says it was made from BF16 weights and can be served with mlx_lm.server as an OpenAI-compatible endpoint.
In the repository description, the author reports roughly 15-16.5 decode tokens per second and about 17 GB peak memory on an M5 Pro with 24 GB of memory. They also report seven passes from eight LeetCode-style smoke tests and a sandboxed terminal-agent run that fixed three bugs in a small repository.
The page does not provide independent benchmarks, reproducible test artifacts, comparisons under controlled settings, or evidence that the reported agent result generalizes beyond the stated small repository.