1.0.2 — Video with sound, an API that stays online, and pasted Hugging Face links

Generate video with synchronized sound, stop a video job, and unload an idle Desktop chat model while the local API stays online. New installs start with a smaller default model, and the CLI and Desktop accept pasted Hugging Face model links.

• Video with sound, and a stop button. Pika H3 Flex generates video with synchronized audio from a prompt, frames, or image, video and audio references. Desktop adds MiniMax H3 and Pika H3 Flex to the Video tab. Stop Generating cancels video jobs on every backend and removes partial output; Desktop keeps the job as “Cancelled”. (#5557, #5559, #5570, #5581, #5564, #5569)
• Idle unload keeps the API online. Desktop unloads an idle chat model while the server keeps running; the next Desktop or local API request reloads it automatically. Settings adds a 5-minute option. Speech models and older runtimes still stop the server when idle. (#5580)

  After the idle timeout	1.0.1	1.0.2
  Local API	stopped	keeps running
  Model memory (Qwen3.5 4B, measured)	freed by stopping the server	3.25 GB → 0.74 GB
  Next request	start the server again	reloads the model automatically

• Faster first run. New installs start with LFM2.5 2.6B, a 1.6 GB download instead of 3.06 GB. Choosing a model that fits goes straight to “Download & start”; only a model this Mac cannot run opens compatibility details. On Macs with 32 GB or more, Qwen 3.5 4B is one click away for image questions. (#5614, #5615, #5626)
• Paste a Hugging Face link. serve, chat, run, pull, rm, info and bench accept huggingface.co, hf.co or hf-mirror.com model URLs and print the org/name they use. rapid-mlx pull <link> --check --json checks whether a model will run without downloading it. Desktop accepts links in model management search or through Add from Hugging Face…; unsupported models say why. Repos with extra per-quantization folders download only the checkpoint that runs. (#5617, #5620, #5621, #5622, #5624, #5627, #5628, #5629)
• A clearer model picker. The chat picker lists downloaded, selected and serving models, with in-progress downloads disabled; “Browse all models…” opens model management. Duplicate aliases are gone, draft-only MTP models stay out of chat choices, and video and catalog models show readable names. (#5566, #5601, #5603, #5609, #5610)
• Simplified Chinese and VoiceOver. More of Settings, Audio, Images, Video and About is translated, including benchmark meter axes and the Vision badge. VoiceOver reads the resident-model lock as “Kept in memory” and skips decorative symbols. (#5540, #5602, #5603, #5611)
• Downloads, MTP, agents and adaptive thinking. pull checks memory and free disk and shows the file plan; a slow-mirror fallback resumes the verified partial file. Eligible models download their pinned MTP head automatically, including on first serve, respecting offline mode and download opt-outs. Agent setup finds the running server, waits during download or loading, supports offline configuration with a named model, and uses the memory-fitted context window. Adaptive thinking uses the server’s bounded reasoning budget, medium (2,048 tokens) by default; an explicit request effort still wins. (#5592, #5545, #5599, #5535, #5582, #5585, #5595, #5604)
• Bug fixes. Memory use matches Activity Monitor instead of treating file cache as unavailable. Failed video jobs stay in history with a reason; saved model selections survive idle release and cancelled catalog refreshes. An unreadable cached model no longer hides the list. Context overflow returns an actionable error, failed processing keeps queued work running, and idle release waits during speculative serving, uploads and Local Search indexing. Interrupted installs stop subprocesses before restoring the previous environment; Apple Silicon detection, upgrade verification and credential redaction in doctor are fixed. (#5614, #5588, #5542, #5543, #5585, #5537, #5544, #5536)
• Upgrade notes. Run rapid-mlx upgrade or curl -fsSL https://rapidmlx.com/install.sh | bash. Existing installs keep their saved model. MTP greedy output can differ from 1.0.1 when two tokens are exactly tied; the target model still verifies every emitted token. Adaptive-thinking requests without an explicit effort use the server’s budget, medium (2,048 tokens) by default. Use pull --no-preflight to stage a model for another Mac. Pika H3 Flex references download about 69 GB on first use; prefetch with rapid-mlx pull pika-h3-flex --with-ref2va. (#5615, #5571, #5604, #5592, #5563)
• Desktop 1.0.2. Published 2026-11-06 as rapid-mac-v1.0.2: video with sound and Stop Generating, idle chat-model unload with the API online, a smaller first-run model for new installs, pasted Hugging Face links, a clearer model picker, and more Simplified Chinese and VoiceOver coverage. Download Desktop 1.0.2. (#5581, #5569, #5580, #5615, #5620, #5601, #5611)