0.16.5 — DeepSeek V4.1 Flash 1017 and steadier streaming

0.16.5 adds an experimental DeepSeek build, keeps interrupted Desktop replies visible, and aligns streamed and buffered API text.

• DeepSeek V4.1 Flash 1017. The experimental mixed-precision alias needs at least 224 GB of memory. It is text-only, greedy-only, and has no tool calling; inputs are capped at 8,192 tokens and outputs at 4,096 tokens. (#4953)
• Desktop chat. Interrupted replies remain visible with Retry, even after relaunch. Chinese chats get Chinese titles and a Chinese assistant identity. Follow-up suggestions use the user's voice. Date questions omit the guess caution; region-only weather questions request a more specific place without a failure card. (#4940, #4929, #4944, #4945)
• Streaming and tools. Split tool-call markup stays out of streamed chat text, and logprobs stay with their own text. Responses and Messages streams keep text that arrives beside a tool call, while buffered replies preserve tool-like text when tools are unavailable. Native-XML models can parse large tool lists under automatic tool choice. (#4950, #4960, #4963, #4927, #4928)
• Uploads and timeouts. Authenticated image and video routes check the API key before reading uploads; video extension uses its own upload limit. Serial TensorFold generation cancels on timeout or shutdown so later requests can proceed. (#4924, #4962, #4949, #4955, #4932)