The Day in AI
In short
- DeepSeek put its chatbot's web search, file input and URL reading into its APIs on 13 November, and added standalone speech and OCR APIs.
- The Endless Frontier released the weights of Fx-Bio-1031, a 180B-parameter bioinformatics agent it says scores 70.59% Pass@5 on BioMysteryBench Difficult.
- MOSI Intelligence and OpenMOSS released MOSS-Live, a full-duplex voice system that answers speech not aimed at it 2% of the time, by its own measure.
- Tencent Hunyuan put Hy Realtime, a voice model that decides on its own when to search the web, into Yuanbao voice calls.
- ggml-org released llama.cpp v1.0.0 and ggml v1.0.0 after about 3.5 years, and both projects now use semantic versioning.
The day in brief
DeepSeek opened its chatbot's web search, file input and URL reading inside its APIs on 13 November, with search billed at $0.75 per 1,000 calls.
Developers turn the features on with parameters in the Messages, Responses and Chat Completions APIs. DeepSeek's servers then run the tool calls and assemble the context, so the developer doesn't write the search loop. The deepseek-flash model, currently serving DeepSeek-V4.1-Flash-1017, supports the integrated features. The release extends the Search and Fetch APIs DeepSeek opened on 17 October.
There are also new standalone APIs for OCR, speech recognition and text-to-speech. DeepSeek charges $0.225 per 1,000 pages for file parsing and OCR, $0.116 per hour of audio for speech recognition and $0.45 per 10,000 characters for speech, and the last two are currently at half price. These tool fees come on top of tokens, which cost $0.30 per million cache-miss input tokens and $1.20 per million output tokens at peak hours. DeepSeek also open-sourced a browser chat client that can import a user's DeepSeek chat history, so anyone can rebuild the app on the API.
The Endless Frontier released full weights for Fx-Bio-1031 under a restricted license. It is post-trained from Alibaba's Qwen3.8-Flash-Next to run multi-step analyses by itself, which means writing scripts, installing software, querying NCBI and Ensembl and checking its own output. The company reports 70.59% Pass@5 on the Difficult subset of BioMysteryBench, which counts a problem as solved if one of five attempts gets it. It puts Opus-5 at 52.94% on the same subset. The lead doesn't hold everywhere, because Opus-5 edges it on the expert-solvable subset and GPT-6-Astra beats it on BixBench3. The model has 6B active parameters per token, and the company pitches it at hospitals and labs that want to run analyses on their own hardware.
Two Chinese voice systems shipped on 13 November. MOSS-Live splits the work between two models. A foreground model listens, handles turn-taking and speaks, and a background model does reasoning and tool calls, so the conversation keeps going while a task runs. MOSI reports that it cuts in during a user's pause 1.44% of the time, against 10.37% for GPT-Live-1. It trails GPT-Live-1 on the τ-Voice multi-turn task score, 57.93% to 67.68%. Tencent's Hy Realtime treats the search decision, the tool call and the spoken answer as one process, and it can sing on request through work with a Tencent Music Entertainment team. Tencent published no benchmark numbers.
Also in the news
- ggml-org will now mark breaking changes to the public libllama and libggml C headers with major-version bumps in llama.cpp and ggml.
- OpenAI launched Daybreak Red Team agents for web app penetration testing and a Defense Factory plugin on 12 November, and the plugin drafts patches that engineers must review.
- OpenAI has disbanded its robotics team as its general models start running robots, according to Fortune.
- Researchers allege that DeepSeek, Kimi and other Chinese models were trained on banned Nvidia Blackwell GPUs, identified with a weight fingerprinting technique, Tom's Hardware reports.
- Nathan Lambert of Interconnects AI argues that Jensen Huang's resignation from Nvidia will very likely end funding that US open-model developers relied on.
- openJiuwen described a plugin system in which WorkSwarm agents write, install and roll back their own extension packages.
- Anthropic's Claude Fable 5.2, released on 9 November, costs $8 per million input tokens and $40 per million output tokens, 20% below version 5.1.
Everything from this day
No records logged on this day.