AI News 2025-12-20
AI Daily Brief
Summary
OpenAI Codex added reusable skill packs feature; Google launched FunctionGemma model optimized for function calling.
Volcano Engine predicts the large model market will grow 10× in 2026; BlueFocus partnered with Doubao model to build an intelligent marketing engine.
Mistral AI released new document recognition technology; Luo Yonghao will release AI software from Thin Red Line Technology.
Lenovo and ByteDance deepened cooperation to preinstall AIGC plugins; Anthropic tested Claude's Task Mode.
Today’s AI News
OpenAI Codex adds “Skills” feature: OpenAI announced that its Codex tool now officially supports “Skills” functionality. This feature allows users to package instructions, scripts, and resources needed to perform specific tasks into reusable skill packs. Users can invoke skills directly through specific commands or let Codex automatically select them based on context. Skills follow an open folder structure and can be installed locally or in team projects. OpenAI also provides some popular out-of-the-box skills, such as generating code from Notion specifications, managing Linear tickets, and automatically fixing GitHub CI errors.
Google releases FunctionGemma model: Google launched FunctionGemma, a new model optimized for function calling. Fine-tuned from the Gemma 3 270M model, it aims to convert natural language instructions into specific commands that can be executed on mobile devices (e.g., creating reminders, adjusting settings), marking a transition from conversation models to executable intelligent agents. On specific test sets, its accuracy improved from 58% before fine-tuning to 85%. The model is suitable for scenarios like developing local phone assistants, smart games, and local AI experiments, and supports offline operation.
Volcano Engine’s Tan Dai predicts 10× large model market growth in 2026: At the 2025 FORCE conference, Volcano Engine President Tan Dai predicted that the large model market will achieve 10× growth in 2026, emphasizing that industry competition should be about growing the market pie together rather than zero-sum games. He pointed out that token usage is a core metric for measuring AI scaled applications. Tan Dai defined 2025 as “the year of the agent” and predicts a full-scale explosion in 2026. Currently, enterprise agents are experiencing dual transitions: from dozens to hundreds or even thousands of deployments, and from entering serious production and high-specialization fields.
BlueFocus announced deep partnership with ByteDance’s Doubao model family: The company will combine Doubao’s text generation and logical reasoning capabilities with Seedance video generation model to build an intelligent marketing engine covering the full chain of “copy-script-video-distribution.” This标志着 domestic large models are accelerating into deep-water commercial applications, with the marketing industry becoming one of the frontiers for implementation.
Mistral AI released new document recognition technology Mistral OCR 3: Its overall performance in tables, scanned documents, and handwritten recognition improved by 74% over the previous generation. The technology can extract text and embedded images with high accuracy, supports Markdown output and HTML-based table reconstruction, with processing costs as low as $1-2 per 1,000 pages, suitable for enterprise scenarios like invoice processing and document digitization.
**Luo Yonghao announced he will focus on releasing AI software developed by Thin Red Line Technology at the 2025 Technology Innovation Sharing Conference in Shanghai on December 30. Conference tickets will be sold in实名制 (real-name) starting December 22; he emphasized this event is about sharing innovation, not livestream e-commerce.
Lenovo and ByteDance are deepening cooperation: The partnership aims to preinstall AIGC plugins on Lenovo and other hardware manufacturers’ AI phones, improving user experience and expanding market entry points. The two sides maintain close communication in the intelligent terminal space, hoping to combine respective advantages to drive AI technology implementation. Industry insiders believe such cooperation could lead the mobile phone market toward a smarter direction.
Anthropic is testing a new feature called “Task Mode” for its AI assistant Claude: The mode uses a dual-panel layout that generates structured “action plans” and proactively asks questions when instructions are unclear. It integrates various skills and tools, can automatically execute complex tasks like code generation and web search, and archives results. The feature is expected to open to the public this month.
ByteDance’s next-generation audio/video creation model “Seedance 1.5 Pro” is now live on the Doubao platform: The model can deeply understand text and simultaneously generate matching visuals, sound effects, and more natural human voice lines, while automatically adding dynamic lens effects. Users only need to upload a reference image and input prompts to one-click generate complete有声 videos (videos with sound), achieving creation from static materials to dynamic content.
WeChat Yuanbao AI assistant added one-sentence reminder feature: Users just need to send a message containing the task and time to Yuanbao, which will automatically recognize and send reminders on time via WeChat service notifications—simple as chatting. The feature supports both text and voice input, especially convenient for elderly users to set daily tasks. Yuanbao, launched in April this year, exists as a WeChat friend without needing to download a separate app. Besides reminders, it also has useful features like parsing official account articles, image recognition, and automatic blurring in ID photos.