0.16.3 — More reliable serving and Local Search

0.16.3 improves recovery under memory pressure, tool-call replies, model loading, and experimental Local Search.

• Serving under load. Concurrent long prompts queue in memory-sized groups. After an out-of-memory batch failure, the server releases its memory for later requests. Nemotron 3.5 text responses and prefix-cache restores remain consistent across mixed-length concurrent prompts.
• Tool replies and MTP. Truncated non-streamed tool calls no longer leak markup in Chat Completions, Messages, or Responses. Undeclared tool calls give a readable unavailable-tool notice across formats, including Nemotron and Gemma 4. Logprobs requests work with MTP speculative decoding.
• Model startup and image runtime. Missing or truncated checkpoint weights stop startup with a cause and recovery step. Pillow is kept below 13 so supported fresh installs can load the image runtime.
• Experimental Local Search. Added folders refresh automatically when Search opens and during background rescans; Whole Mac refreshes on the background interval. Meaning-based searches with no match show up to three closest files. Folder access denials offer Privacy Settings and Try Again, and Quick Look previews no longer crash Desktop.
• Desktop. Opening What's New after an update and the Star on GitHub prompt no longer quit the app.