0.13.3 — GLM-5.3-Flash, and vision chats that stop recomputing

Released 2026-09-01 · full changelog · GitHub releases

Upgrade: pip install -U rapid-mlx · brew upgrade rapid-mlx · or grab the desktop app.

A 320B-parameter model now answers on a desktop: glm5.3-flash-4bit runs through the CLI, the OpenAI-compatible server and the Mac app. Vision models stop recomputing the text they have already seen, finished video jobs survive a restart, and the app gains a command palette.

GLM-5.3-Flash is served. rapid-mlx serve glm5.3-flash-4bit runs Zhipu's 320B MoE (18B active per token) with the processor and runtime compatibility the checkpoint needs. Measured on a 256 GB M3 Ultra: a sustained 512-token response decoded at a median 29.2 tok/s using 165.4 GB of active MLX memory, so the alias asks for the 192 GB tier. Speculative decoding is deliberately off for it — its qualification run showed no throughput gain, and we would rather ship the honest default than a flag that looks impressive.
Vision chats reuse the text they have already processed. A hybrid multimodal conversation whose current turn is text-only can now reuse its text prefix instead of recomputing it, with cache identity bound to the rendered request and model state. Requests that actually carry media still go through the validated vision path — a picture never scores a text-only cache hit.
Finished video jobs survive a restart. Completed generations persist across server restarts, and a client can ask what a model is capable of before paying to load it.
A command palette in the Mac app. Common actions are reachable from the keyboard, a diagnostics shortcut makes failures easier to report, model pickers stay scoped to the task you are in, and streaming Markdown keeps rendering when a window comes back from a disconnected display.
Packaged builds can repair themselves. A Desktop install that fails after the DMG step can now recover its engine instead of leaving a half-installed app.