腾讯研究院AI速递 20260106

Group 1: Notion 3.0 Update - Notion 3.0 introduces personalized agent customization, integrating the latest models GPT-5 and Claude Sonnet 4, expanding the MCP list to include applications like Lovable and Perplexity [1] - Users can now complete complex tasks such as database setup and automation using natural language, significantly lowering the usage barrier for previously deterred users [1] - The commercial subscription is priced at $24 per month or $20 per month when billed annually, with a 14-day trial period, targeting users needing personal finance and data management systems [1] Group 2: Kiwi-do Model Emergence - A mysterious model named Kiwi-do has emerged in the large model arena, claiming to be from Kimi, with training data up to January 2025 and passing the VPCT visual physical understanding test [2] - Speculations suggest Kiwi-do may be related to the K2-VL multimodal version mentioned in AMA activities or the new K2.1/K2.5 models expected to launch in Q1 of this year [2] - Kimi has completed a C-round financing of 3.5 billion yuan, with cash reserves reaching 10 billion yuan, potentially accelerating the training of the K3 model with aggressive GPU expansion [2] Group 3: WeChat AI Mini Program Growth Plan - WeChat mini programs have launched an "AI Applications and Online Tools Mini Program Growth Plan," providing comprehensive support including cloud development, AI computing power, data analysis, and commercial monetization throughout 2026 [3] - Developers can receive a free personal cloud development environment for six months, 100 million tokens for Tencent's mixed yuan 2.0, 10,000 image generation credits, and a one-year subscription to WeAnalysis Pro [3] - The platform will enable virtual payment and membership subscription capabilities across all terminals with limited-time discounted rates, with AI mini programs like "Guess Salt" and "Style Converter" already showing commercial potential [3] Group 4: Huawei's Open Source Model - Huawei has released the open-source multimodal model openPangu-VL-7B, utilizing Ascend's native architecture, achieving a first-word inference delay of only 160 milliseconds on a single Atlas 800T A2 card at 720P, enabling real-time inference at 5 FPS [4] - The model completed stable training on over 3 trillion tokens during the pre-training phase, using a high-performance visual encoder adapted for Ascend, improving throughput by 15% compared to the ViT-H series [4] - It excels in tasks such as general visual question answering, document chart understanding OCR, visual localization, and short video comprehension, using relative coordinates for positioning [4] Group 5: Samsung's AI Home Ecosystem - Samsung unveiled its AI home ecosystem at CES 2026, featuring a 130-inch Micro RGB TV equipped with Vision AI Companion (VAC) for recommending movies, recipes, and music, along with an AI football mode for realistic match experiences [5] - The AI refrigerator Family Hub, powered by Google Gemini 3, automatically tracks ingredient usage and recognizes specific sounds to provide personalized dietary reports, supporting seamless integration with connected kitchen appliances for automated cooking processes [6] - Samsung plans to increase the number of mobile devices equipped with Gemini AI features to 800 million by 2026, with a health companion function that aggregates data from connected devices to proactively alert users to abnormal health signs and share data via the Xealth platform [6] Group 6: ima's PPT Generation Feature - The ima 2.1.3 update introduces a PPT generation feature that automatically creates charts and highlights key points from user data, supporting various styles [7] - This feature is designed for academic reports, workplace summaries, and proposal promotions, addressing challenges of time constraints and content distillation, simplifying the process of creating year-end summaries and final reports [7] - The new version is now live, allowing users to update and experience the enhancements, as ima continues to iterate and upgrade its products in response to user needs [7] Group 7: Chinese Eyewear Brands at CES 2026 - At CES 2026, 16 out of 23 AI eyewear exhibitors are Chinese brands, including Alibaba, Thunderobot, and Rokid, with a focus on brand exposure and international expansion [8] - The budget for independent exhibition booths starts at 1 million yuan, with TCL having the largest exhibition area, slightly surpassing LG and Sony [8] - Various Chinese hardware brands, including humanoid robots and AI companion robots, are showcased, while the density of Chinese smartphone manufacturers is lower as they shift focus to the MWC exhibition [8] Group 8: ByteDance's SeedFold Model - ByteDance's Seed team has proposed the SeedFold molecular structure prediction model, achieving state-of-the-art results in the FoldBench benchmark, surpassing AlphaFold3 through width expansion and large-scale data distillation [9] - The efficient variant SeedFold-Linear employs a linear triangular attention mechanism, reducing computational complexity from cubic to quadratic, and constructs a large-scale distilled dataset containing 26.5 million samples, expanding experimental structural data by 147 times [9] - Experiments indicate that the folding model's capacity is primarily limited by the paired representation hidden dimension (128 dimensions), with SeedFold outperforming AlphaFold3 in antibody-antigen predictions and SeedFold-Linear excelling in protein-ligand predictions [9] Group 9: Midjourney Founder Insights - The founder of Midjourney, David, shared that he completed more programming projects during the Christmas holiday than in the past ten years combined, with notable comments from Elon Musk about entering a singularity [10] - Engineers from Anthropic noted that programming agents, particularly Claude Opus, can compress six years of work into a few months, with Google's chief engineer generating a distributed agent coordinator in a day that took a year to build [10] - LiveBench's latest evaluation ranks Claude 4.5 Opus at the top, with developer Boris Cherny sharing his setup that achieves 50-100 pull requests weekly [11]