Changelog / release

1.0.4 — Reliable agent launches, image edits and local adapters

Released 2026-11-17 · full changelog · GitHub releases

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

Rapid-MLX 1.0.4 makes coding-agent launches reliable, adds MiniMax-H3-Image and FLUX 3 Klein, and serves local LoRA adapters without fusing them.

  • Coding agents on first launch. Setup configures the agent before waiting for server readiness, with bounded waits and clear messages. Claude Code uses the real context window without a billing notice; removal restores prior settings while keeping later edits. Hermes preserves the first message and cleans up with --remove. Aider uses the user’s shell and isolated model metadata; Qwen Code gets valid empty settings in its isolated home. (#5804, #5811, #5840, #5813, #5815, #5828, #5818, #5812)
  • Image generation and multi-reference edits. minimax-h3-image (4-bit) and minimax-h3-image-8bit support up to 16 API reference images; Desktop currently limits attachments to ten. The MiniMax-H3-Image Community License excludes the EU, UK, Republic of Korea and US, and requires separate authorization above US$1M annual revenue or for hosted third-party services. Responses carry attribution; weights download from the upstream Hub only. FLUX 3 Klein and Klein Base support up to ten edit references. Qwen-Image advertises checkpoint-specific limits and returns HTTP 400 when exceeded. (#5832, #5790, #5839)
  • Image LoRAs. Place your .safetensors files in ~/.rapid-mlx/loras and select them by name for generation and edits. The image runtime moves to mflux 0.23 with no generation behavior change. (#5792, #5787)
  • Local text adapters. rapid-mlx serve <model> --adapter <dir> applies an mlx_lm.lora adapter at startup without fusing or re-quantizing. Base-model checks work with aliases, Hugging Face repo IDs and local cache paths; mismatched bases are rejected. (#5808, #5842)
  • API and built-in tools. /v1/completions returns prompt-token logprobs with echo and logprobs. Qwen3-Reranker uses its trained prompt format. Lone UTF-16 surrogates in embedding or rerank input return HTTP 400. The built-in calculator and batch read-only helpers accept native JSON arguments. (#5809, #5843, #5801, #5800)
  • Cached first-token latency. Single requests hitting long prefix caches avoid extra KV copies; chat prompts are tokenized once per request. Decode speed and outputs are unchanged. (#5831, #5803, #5810, #5829)
  • Desktop fixes. Conversation history remains reachable from the compact sidebar rail, and onboarding keeps model names visible beside long size details. (#5799, #5793)
  • License detection. The project LICENSE is the verbatim Apache-2.0 text. (#5783)
  • Upgrade and Desktop 1.0.4. CLI v1.0.4 and Desktop rapid-mac-v1.0.4 were released 2026-11-17 (PT). Run rapid-mlx upgrade, brew upgrade rapid-mlx or pip install -U rapid-mlx==1.0.4. Download Desktop 1.0.4.