0.16.9 — Smoother Desktop chats and reliable model loading

Browse models while Desktop replies stream, keep your chat instructions intact, and reuse long prompts more efficiently. Model downloads now follow validated revisions, with clearer recovery when setup needs attention.

• Desktop feels steadier. The model picker stays responsive during a streamed reply, and switching models still asks before interrupting it. Desktop verifies its bundled engine before launch, preserves model caches through full-app upgrades, and offers to connect a coding tool after your first successful chat. The newsletter is available in About and at rapidmlx.com/newsletter. (#5176, #5164, #5243)
• More predictable chats. Gemma replies stop at their turn terminator; custom system instructions remain verbatim when tools are offered automatically, while required tool constraints remain enforced. (#5185, #5235)
• Repeated long prompts. Hybrid cache keeps useful restored boundaries under cache pressure. In the engine's Qwen3.6-35B workload with approximately 32K-token prompts and a 2 GiB prefix budget, median repeat-request time to first token fell from 18.622 s to 0.390 s on M3 Ultra 256 GB. This workload-specific measurement is not a general speed guarantee, and cache residency can change output tokens. (#4459)
• Model loading and recovery. Catalog models use validated checkpoint revisions; a stale cache downloads the pinned revision before serving, while a pinned cache works offline. Explicit pulls verify content before reporting success. Two Nemotron 3.5 Bolt E4B text aliases add 4-bit and 8-bit options with tool calls and thinking controls; image input and draft acceleration are not included. Missing required dependencies show the repair command and run it only with interactive consent. (#5236, #5154, #5138, #5242)
• Serving controls. Qualified Qwen3.8-27B and Nemotron profiles can opt into --tensorfold-lanes {1,2,4,8}; DeepSeek V4.1 uses --dsv41-lanes {1,2,4,8}. Both remain serial by default. Unset the retired RAPID_MLX_DSV41_LANES variable before using the CLI option. (#5231, #5230)