Models on Apple Silicon

A series of open releases, published as VisualInference on Hugging Face, that adapt large open-weight models to Apple’s MLX framework.

Every number below was measured on one machine — a Mac mini with an M4 Pro and 48 GB of memory — and each release documents its harness and its limits alongside the result.

Ideogram 4 (9.3B): I folded classifier-free guidance into a single branch, then distilled the 12-step schedule down to 6. At 800×992, generation drops from about 250 seconds to about 86 — 2.9× faster end to end, with no refusal placards in my test set.

Qwen3.8-27B (AEON variant): I published a 6-bit MLX quantization and split the model’s own multi-token-prediction head into a drafter for speculative decoding. On coding prompts, decode rises from 11.4 to 21.5 tokens per second.

An agent lane: a serving recipe, a thinking-budget proxy and four patches for running a 27B Qwen3.8 as a local agent backend — 15–19 tokens per second end to end at 24–27k tokens of context.