0.13.4 — Qualified Qwen models go faster on their own, and benchmarks stay on your Mac

Released 2026-09-03 · full changelog · GitHub releases

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

Four qualified Qwen artifacts now pick their validated speculative-decoding preset without being asked, so concurrent work finishes sooner with no flag to remember. Video generation gets a real workspace in the Mac app, the Community Benchmark keeps every result local until you explicitly share one, and a model that will not fit now says so before it tries to load.

Qualified Qwen models select their own MTP preset. Qwen3.5 4B/9B, Qwen3.6 27B and Qwen3.8 27B (the exact qualified 4-bit artifacts) now choose their validated multi-token-prediction preset and the continuous scheduler when no speculative-decoding option is given. Across mixed four-request cohorts, aggregate throughput improved 14.1%–30.8%. For Qwen3.8 27B the single-request path also scales with context: 1.43× the 0.13.3 ordinary path at 128 tokens and 2.34× at 32K. --no-spec-decode restores ordinary decoding; Desktop has the same switch.
Video generation in the Mac app. Enable it under Experimental Features for a local workspace with queued jobs, progress, cancellation and restart-safe results. The signed app carries the runtime and encoder for LTX 2.5 q8 (text and image to video), Wan 2.1 1.3B bf16 and CogVideoX-Fun 5B q4 — nothing to provision after install. Wan 2.1 needs a Mac with at least 40 GB of unified memory.
Community Benchmark is private by default. rapid-mlx benchmark and the new Desktop page plan, run, validate and archive text, image and video measurements locally. Nothing leaves the Mac unless you choose to send it, and the consent step shows the exact destination and payload digest before you do.
A model that will not fit tells you first. CLI, server and Desktop now read one atomic model registry, per-model Metal limits are derived from the machine's available unified memory, and a rejected launch returns fit guidance instead of failing deep inside model loading.
Headless deployment is documented and qualified. A LaunchDaemon guide covers dedicated service accounts, loopback-first networking, KeepAlive, reboot recovery and FileVault — see Run Rapid-MLX at boot on a headless Mac.
Desktop quality of life. Opt-in persistent memory across conversations (editable and removable in Settings), Mermaid blocks rendered as isolated network-blocked previews, dictation that overlaps model page-in with your speech, and an update card that says whether the installed app is current, behind, or ahead of the public release.