mlx-optiq ships early Gemma 4, Qwen quants
mlx-optiq appears to be a legitimate Apple Silicon MLX quantization project: it has a live PyPI package, a dedicated docs site, and Hugging Face model cards for the new OptiQ releases. The specific Qwen3.6 and Gemma 4 OptiQ checkpoints are real uploads from mlx-community, but they are language-only variants by default, with vision/audio stripped out unless you re-convert the upstream base model. Public validation so far is mostly the maintainer’s own model cards plus a few community threads and adjacent MLX benchmarks, not broad independent bake-offs yet.
Hot take: legit project, but treat these as promising early releases rather than battle-tested defaults.
- –`mlx-optiq` has a real package and docs footprint, which is a strong legitimacy signal.
- –The HF cards for `Qwen3.6-27B-OptiQ-4bit`, `Qwen3.6-35B-A3B-OptiQ-4bit`, `gemma-4-26B-A4B-it-OptiQ-4bit`, and `gemma-4-31B-it-OptiQ-4bit` all describe sensitivity-aware mixed-precision quantization.
- –The cards explicitly say the OptiQ builds ship the language stack only; the multimodal/vision parts are dropped by default.
- –There is some early community testing around nearby Qwen 3.6 and Gemma 4 MLX quants, but I did not find widely cited independent benchmarks specifically for these OptiQ uploads yet.
- –Practical read: likely safe to try if you want Apple Silicon serving/tuning, but I would not assume quality or stability parity with the upstream base models until more users report results.
DISCOVERED
4h ago
2026-04-27
PUBLISHED
5h ago
2026-04-27
RELEVANCE
AUTHOR
Intelligent_Ice_113