raullenchai/rapid-mlx released 1.0.2. The November 6, 2026 commit f1aa011 only bumps the version from 1.0.1 to 1.0.2, and the contents are summarized in the CHANGELOG and docs/release-notes/v1.0.2.md. The repository description states that this repository is an Apache-2.0 open-source inference server running on Apple Silicon, that it provides OpenAI- and Anthropic-compatible APIs, and that it covers tool calling for coding agents.
Most of the changes in this version are on the desktop app side. Audio and reference (first/last-frame, reference) controls were added to video generation, and the target models are listed as MiniMax H3 and Pika H3 Flex. Video jobs in progress can now be interrupted. It is stated that the tool that decodes video and audio references is included in the app, so no separate installation is needed. A method that unloads only the idle chat model while keeping the server running was also added. Options in 5-minute increments appeared, and the sidebar displays the unloading and reloading states. Included along with these are a fast first run that starts with LFM2.5 2.6B (1.6 GB) on a new installation, adding models by pasting Hugging Face links, Simplified Chinese and VoiceOver accessibility improvements, and tidying up the chat model selection list. The commit stats are 110 lines (101 added, 9 deleted), and the Info.plist version marking changed from 1.0.1 to 1.0.2, and the build from 195 to 196.
The comparison target is the desktop app release. The files touched by this commit are four: CHANGELOG, Info.plist, docs/release-notes/v1.0.2.md, and pyproject.toml, which are changes to the records and version markings rather than to code. Video generation interruption works only when there is a server that supports cancellation, and idle release still stops the server on older runtimes and voice models. The repository description's claim of up to 4x speed compared with mlx-lm, and the community benchmark results, were not reproduced in this commit and no figures are given. It is premature to judge what 1.0.2 actually improved until the desktop app is checked directly.