Model families on rapid-mlx
228 MLX-native models across 15 vendor LLM families plus Audio and
Video — 175 text aliases (chat,
vision & reasoning) across the vendor LLM families, a 44-alias Audio family covering
OpenAI-compatible TTS and STT, and an 9-alias Video generation family
behind the async /v1/videos API. Tier-1 support: Qwen 3.6,
Gemma 4, DeepSeek, and GPT-OSS. Each family page
groups its version lines (e.g. Qwen3.5 / 3.6 / Coder / VL, or TTS /
STT for Audio) under one consolidated deep dive with the alias list,
the tool-call parser, reasoning parser, and capability notes. Six
models also get a hero deep dive because the engineering work to
make them serve cleanly on MLX is non-trivial, and rapid-mlx shipped
that work first or shipped it best.
Where to look
aliases.json grouped by family — a flatter list, easier to ctrl-F.Hero deep dives
Six models get a per-page deep dive because the engineering work to make them serve cleanly on MLX is non-trivial, and rapid-mlx shipped that work first. These sit outside the family pages — they are their own hand-written stories.
qwen3_xml tool calling, hybrid-cache fix, vision-tower strip.mlx-lm still reports it as unsupported.ui_tars tool parser.reasoning_content on /v1/responses.holo3.1-35b-a3b (4-bit / 8-bit).All families
Every alias in vllm_mlx/aliases.json belongs to one of
the 14 vendor LLM families below; the audio aliases (from
vllm_mlx/audio/aliases.json) live on the 15th —
Audio and Video generation. Each family
page nests its version lines under <h2> headings
so all the alias / parser / variant detail is one click away — no
per-version page proliferation. Pass any alias to
rapid-mlx serve <alias>; the plumbing gets wired
up for you.
llama tool envelope.muse ATEM tool envelope, text serving today.glm47 tool envelope.hy_v3 tool + reasoning parser.lfm tool parser./v1/audio/*./v1/videos job API · needs Python 3.11+ and ffmpeg.Notable small models
Headline-family pages are organized by the big lines (Qwen, Gemma, Llama). But the most interesting work in 2026 is often happening at the small-and-weird end — focused-purpose models, single-shop research checkpoints, frontier-style tricks compressed into 4 GB. The list below is editorial: models we've watched ship, models another team has reported working well in production, or models on our short-list for the next 0.9.x release.
holo3.1-35b-a3b (4-bit / 8-bit).reasoning_content on /v1/responses.qwen3_xml tools + a custom hybrid-cache fix.mlx-lm still reports unsupported./v1/embeddings for RAG without an API call.Which RAM tier?
For "what fits on my Mac" the easiest path is the live picker at models.rapidmlx.com. The landing page also has a static summary table per RAM tier (see "Which model fits your Mac?").