0.16.6 — One-command agent launches and Qwen-Image 2.1

Launch a coding agent against a private local profile, return to the same connection after a restart, and reuse a cached or already served model. Qwen-Image 2.1 becomes the default image alias.

• Coding agents. rapid-mlx launch <agent> keeps its own profile, loopback key and port across restarts. Codex, Claude Code, OpenCode, Pi, Hermes and Crush can use this path without changing the user's existing agent configuration. A second agent reuses a compatible model already served locally; a complete cached chat model can start without a download. Claude Desktop connects through its official in-app Gateway. Claude Code authentication, Codex model selection and OpenCode 2.x launch have fixes. (#4979, #4981, #4998, #5016, #5026)
• Image generation. qwen-image now selects the Qwen-Image 2.1 4-bit build for generation and editing on an 8 GB Mac. New qwen-image-2.1-turbo uses the published eight-step schedule for text-to-image generation on a 32 GB Mac. Concurrent image requests no longer fail with a worker-thread error. Existing qwen-image and flux-schnell caches download their repinned weights once; an existing qwen-image-2.1 cache can be reused. (#5000, #4993, #5003, #5005)
• DeepSeek V4.1 Flash 1017. The default revision moves to a new GPTQ build and re-downloads once. Single-stream decoding is faster, rejected speculative verifies preserve output, load-time file caches are released sooner, and concurrent requests can opt into batched decode. The release notes describe a 256 GB Mac requirement. (#5036, #5024, #4987, #5034, #4982)
• CLI recovery. Starting a complete cached chat model avoids a new download, failed pulls and startup report a cause and retry command, and pull and rm ignore models in external folders. (#4971, #4965, #5031)

Try the one-command local agent launch: Codex, Claude Code, OpenCode, Pi, Hermes, Crush, or Claude Desktop Gateway.