0.16.4 — Simpler Desktop chat and new server APIs

0.16.4 simplifies Desktop chat, adds document reranking and cache prefill, and expands the model catalog.

• Desktop chat. Personal Intelligence has been removed; ordinary Chat offers Web Search, Browse Web Page, Weather, Read Attached Document, and configured connectors. Local file and command tools are no longer in Chat. Unreadable tool requests retry once, tool-budget replies give an answer or clear fallback, and Qwen XML tool calls no longer leave empty replies or raw markup. Plain writing and coding turns trigger fewer unrelated web calls. (#4860, #4865, #4876, #4879, #4885)
• Messages and attachments. Chat accepts Word .docx text, and pasted files attach instead of inserting a path. Drafts persist per conversation. Send starts a downloaded but stopped model and keeps the draft if startup fails. The starter assistant now uses a general instruction, with generated titles and follow-up suggestions. (#4895, #4867, #4866, #4878, #4903, #4908)
• Weather and browsing. Weather handles daily forecasts, state and province abbreviations, country aliases, and more Chinese city and province names. Region-only requests prompt for a city. Same-site redirects no longer ask for page-fetch approval twice. (#4893, #4898, #4915, #4921, #4919)
• Retrieval and caching APIs. POST /v1/rerank scores documents with a configured rerank model; requests have bounded document and embedding input defaults. POST /v1/cache/prefill warms a fixed prefix or revisioned draft without returning generated text. (#4832, #4839, #4884)
• Serving behavior. Image routes now require the configured API key. Responses honors max_tokens as an alias of max_output_tokens. Oversized uploads return a 413 error, and /healthz reflects idle model unload. Tool-call examples inside fenced code stay visible as text. (#4851, #4815, #4831, #4834, #4775)
• Models and performance. Experimental MiniCPM-V 4.7 vision aliases and LFM2.5-24B-A2B join the registry. Long-prompt prefill admits concurrent requests when memory allows; multimodal Qwen MoE text decoding and opt-in MTP cohorts run more efficiently. (#4745, #4852, #4835, #4825, #4822)