0.16.2 — More local models, smoother serving and Desktop controls

0.16.2 adds Nemotron 3.5 and small Bonsai 2 models, improves concurrent text serving and agent setup, and expands Desktop controls.

• Models. Nemotron 3.5 Bolt 74B-A9B and Super VL 120B-A12B have 3-bit and 4-bit text aliases for chat, reasoning and tools; these checkpoints do not accept image or video input. Small Bonsai 2 aliases at 2B, 4B and 9B support image input, tools and reasoning.
• Serving. Qualified Qwen3.5 vision servers batch text requests, eligible long prefills yield to other requests, and the experimental Qwen27 text runtime reuses warm prefixes at message boundaries. Qwen3.6 27B recovers after larger concurrent bursts. Speculative decoding falls back for unsupported targets or constrained requests; continuous MTP accepts depth 1 or 2 and clamps higher values to 2.
• Tool responses and recovery. Truncated tool calls report the protocol's length or incomplete reason, forced calls no longer expose partial headers, and undeclared tool calls give a useful reply. Invalid image parameters return a validation error. Interrupted pulls and startup failures give clearer recovery steps.
• Agent setup. Cline, Goose and Crush join the local setup commands; Continue and Kilo guidance is refreshed. Interactive chat-server startup prints usable request and setup commands.
• Audio and long prompts. The speech API's speed setting changes playback duration for Qwen3-TTS, Chatterbox and IndexTTS. Opt-in --pflash-threshold-basis uncached compares the uncached remainder against the PFlash auto threshold when the memory-aware prefix cache applies; total remains the default.
• Desktop and sharing. Settings → Appearance → Text size enlarges interface text. Experimental Local Search is off by default and can be enabled for chosen folders under Settings → Experimental. Share Compute remembers its provider account and releases models on Stop. Plain shares require a current client: clients older than 0.16.2 cannot register with the updated relay.