Changelog / release

1.0.5 — Live monitoring, Raon-1 nano and persistent Desktop images

Released 2026-11-19 · full changelog · GitHub releases

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

Rapid-MLX 1.0.5 adds live monitoring and native Raon-1 nano, improves qualified decoding paths, and makes Desktop images and interrupted downloads survive relaunch.

  • Monitoring and integration. rapid-mlx top shows server activity, model memory, queue depth and cache usage. --once or redirected output takes one snapshot without prompt or response text. rapid-mlx ps --json exposes versioned server metadata without credentials; rapid-mlx ls --json lists cached models. Trusted-host rejections return actionable HTTP 400 guidance. The Build on Rapid-MLX guide covers background servers, model management and monitoring contracts. (#5891, #5894, #5892)
  • Desktop resource card. The menu bar shows resident models, memory, active work and generation throughput, with guarded per-model unload and whole-pool release. (#5893)
  • Raon-1 nano. rapid-mlx serve raon-1-nano uses the bundled native implementation without executing checkpoint code. The official 4-bit model fits 8–16 GB Macs and has dedicated reasoning and tool-call parsing. Streamed arguments preserve literal closing tags inside values. Readiness recognizes the bundled runtime; external checkpoint code still requires consent. (#5861, #5876, #5877)
  • Qualified decode and short-turn tuning. Qwen3.8-27B 4-bit MTP code edits verify longer prompt copies on M5 Max with 36 GB; short single-request MTP turns avoid unnecessary prefill splits and admission waits. Gemma 4 26B-A4B gains eligible single-stream fused expert decoding with a checked fallback. LFM2 MoE fused experts are enabled only on M5-generation GPUs; other or unknown hardware keeps the existing path. (#5852, #5862, #5864, #5873, #5874, #5888)
  • Downloads and startup. Interrupted downloads retain partial data and offer Resume download; source switches reuse verified partial data and temporary upstream failures get bounded retries. FLUX 3 Klein downloads and disk checks include only runtime components. Fully cached pinned drafters can load offline without revision references; incomplete snapshots explain the failure. Failed starts record classified reasons and next steps. (#5841, #5853, #5849, #5854)
  • Desktop images and first run. Generated and edited images persist across relaunch, newest first; deleting one removes its file. Images-view downloads show live progress, bytes and speed. The compact sidebar says Agent, and connections honor the configured home directory. FLUX.2 Klein 4B and Z-Image Turbo are admitted on 16 GB Macs using catalog minimums; FLUX 3 Klein remains unavailable there. The terminal installer offers a first chat only in interactive sessions; internal commands no longer consume the first-use disclosure or count as app opens. (#5856, #5855, #5857, #5846, #5845)
  • Tool-call recovery. Non-streaming chat retries required multi-tool answers with no calls and schema-invalid required or pinned calls once with guided decoding. Successful recovery combines usage; unavailable or unsuccessful recovery preserves the original error. Strict Responses reuse the tool parse. Parsing removes duplicate closers and preserves separate parallel streamed calls, including repeated calls to one tool; an unquoted value ending in a literal closer can still lose that tag, while quoted values preserve it. (#5879, #5895, #5866, #5867)
  • Embedding discovery and cache retention. /v1/models and /v1/models/{id} advertise known embedding dimensions, supported dimensions, input limits, pooling and normalization. POST /v1/cache/clear accepts keep_recent to release older text prefixes while retaining recent entries and protecting active or pinned state. Responses and cache statistics report released and retained state; no-body full clearing is unchanged. (#5889, #5890)
  • Browser extensions and computer use. --cors-origins accepts chrome-extension://*, moz-extension://* and safari-web-extension://*; each permits all valid extension origins for that scheme. Cross-origin access stays closed by default. Computer use fills segmented date/time fields with validated keystrokes, reads empty segments as empty and reports neighboring changes as unverified. (#5878, #5863)
  • Serving correctness. Disabling speculation sizes the context cap from actual cache requirements and rejects over-cap text prompts before prefill. MTP sampling uses the plain-serving filtering distribution, including seeded requests. Unset optional LFM2 layer-layout fields no longer fail admission. Qwen3-Reranker and local-adapter base-model resolution fixes are included. (#5859, #5871, #5875, #5843, #5842)
  • Upgrade and Desktop 1.0.5. CLI v1.0.5 and Desktop rapid-mac-v1.0.5 were released 2026-11-19 (PT). Run rapid-mlx upgrade, brew upgrade rapid-mlx or pip install -U rapid-mlx==1.0.5. Download Desktop 1.0.5.