1.0.1 — Python 3.12, a memory check before download, and five speech models

Rapid-MLX now requires Python 3.12 or newer. On Python 3.10 or 3.11, pip install -U rapid-mlx and older rapid-mlx upgrade report success but leave you on 1.0.0. Move to Python 3.12 the way you installed: rerun curl -fsSL https://rapidmlx.com/install.sh | bash (keeps your settings, models and extras), uv tool upgrade --python 3.12 rapid-mlx, or pipx reinstall --python python3.12 rapid-mlx; with plain pip, install into a new Python 3.12 virtual environment. Homebrew and the Desktop app need nothing. (#5379, #5420)

• Know before you download. rapid-mlx serve checks whether a catalog model fits this Mac's GPU memory before downloading it, and suggests a smaller model instead of failing at startup. (#5402)
• Five new speech models. Speech-to-text adds granite-speech, firered-asr2 and nemotron-asr; text-to-speech adds voxcpm2 and moss-tts-nano. The new speech-to-text models also appear in Desktop's Speech to Text settings. (#5381)
• Benchmark what serve runs. rapid-mlx benchmark run uses serve's default speed settings, accepts serve's acceleration flags, measures MTP and DFlash runtimes and records the configuration that ran; --baseline keeps the original all-off protocol. (#5376, #5382, #5389, #5392, #5397, #5411)
• Fewer stalls and re-downloads. Strict JSON-schema requests can no longer emit whitespace for minutes and hold up coding-agent turns; Qwen3.8-27B models with a built-in MTP head start from an external folder without re-downloading; rapid-mlx share --install-service runs the install that created it. (#5373, #5396, #5370)
• Upgrades that tell the truth. rapid-mlx upgrade checks the installed version in a fresh process before reporting success; if the latest release needs newer Python, it installs the newest compatible release and prints the migration command. (#5421, #5491)
• Downloads you can stop. Ctrl-C cancels a model download within seconds and running the same pull again continues from the files already cached. A model refused for memory now suggests one already downloaded, or one of the same kind that fits. (#5523, #5519, #5513)
• Desktop 1.0.1. Speech to Text settings download other speech models in place with steadier progress; first run, the Images tab, Share Compute and the agent page are fully translated in Simplified Chinese; download failures explain the cause and offer "Switch source"; model sizes and names are consistent everywhere; VoiceOver reads the memory meter and dictation switch. (#5374, #5377, #5522, #5526, #5465, #5433, #5416)
• Desktop anonymous usage data. The anonymous first-run funnel now reports why a model download or first engine start failed, as one cause from a fixed list plus the download source. No error text, paths or model names are sent; see what is collected. (#5404)