Learn more about Claude and ChatGPT's updated news and information about large models. This blog focuses on tracking and parsing the current state-of-the-art Large Language Models (LLMs) trends, technology updates, and their practical applications in different domains.
阶跃星辰(StepFun)团队开源了NextStep-1，一款14B参数的纯自回归图像生成模型。该模型直接在连续视觉空间生成图像，无需依赖扩散模型或离散化处理，由14B参数Transformer骨干和157M参数流匹配头组成。它支持高保真文生图及精准图像编辑（如物体增删、背景修改），在GenEval（0.73）、GenAI-Bench等基准测试中表现优异，接近顶尖扩散模型。但存在生成不稳定、解码延迟等挑战，标志着自回归图像生成新阶段。
Nanobrowser是近期在GitHub爆火的开源AI浏览器自动化框架，上线一周获17,000+星标。其核心采用双智能体协作模式：Planner拆解自然语言指令为操作步骤，Navigator在真实网页中执行、读取等操作。该项目支持本地运行及多模型接入，可实现论文抓取、比价、舆情监控等网页自动化任务，典型案例显示其2分半完成论文数据抓取，成本仅0.1元。
Web3已从概念走向现实，2025年全球市场规模达213.5亿美元，中国相关产业规模超200亿元。其核心在于用户主权，通过区块链、智能合约、NFT和DID实现权力重分配。五大应用场景包括：DeFi（TVL超1200亿美元）、NFT实用化（如星巴克权益）、DAO（活跃组织超5000个）、GameFi（超3100款游戏）和去中心化身份。市场正从投机转向价值驱动，未来机会聚焦创作者经济、数字身份及RWA资产代币化，目标重建数字信任与公平。
LTX-2是Lightricks发布的全球首个音画同步4K视频生成模型，可生成20秒、50fps高清视频，支持文本/图像输入。它实现了角色口型与语音同步，能在ComfyUI运行并本地部署，将于5年11月下旬开源。作为专业级创作工具，LTX-2让"文字变电影级短片"成为现实。
Blockchain, Bitcoin, Web3 in 2025 has made it clear that the price of "digital gold" exceeded 110,000 U.S. dollars, with an all-time high of 111,013 U.S. dollars; blockchain has become a "new infrastructure" and is used in government affairs, finance and other fields, and the RWA market size has reached 202.5 billion U.S. dollars; Web3 market size has reached 21.35 billion U.S. dollars, and is expected to reach 5.1 trillion U.S. dollars in 2030. The market size of Web3 reaches $21.35 billion, turning to real applications, and is expected to reach $5.1 trillion in 2030. China supports blockchain but focuses on Web3 "coinless" path.
Cursor 2.0 is officially released with the launch of Composer, a self-developed large model that generates code at a speed of up to 250 tokens/second, two times faster than GPT-5 and Claude Sonnet 4.5. The model is specially trained for real development scenarios and can autonomously complete the whole process of coding, testing and bug fixing, and is currently only integrated into the Cursor editor for internal use.
FlowithOS is the world's first native operating system designed for AI intelligences, capable of performing complex tasks rather than being limited to chatting. Based on the Chromium browser, the system supports cross-platform multitasking in parallel with unlimited contextual memory and Skills library, and can automate real-world tasks such as Taobao shopping, microblogging and data collection. Currently in early beta, it supports Windows and macOS, and requires an invitation code to experience.
MiniMax released a new generation of open source big model M2, with the performance among the top 5 in the world, and the price is only 8% of Claude 4.5. The model has a total parameter of 230B, an activation parameter of only 10B, and an inference speed of more than 100 tokens/second. It has excellent performance in programming, Agent workflow and multimodal tasks, breaking the "impossible triangle" of high performance, low price and high speed in the field of AI.
In the AlphaArena AI Coin Speculation Live Arena, six top AI models each traded independently in the cryptocurrency market with $10,000 in real money. As of Oct. 23, Qwen3 Max (Alibaba) is in first place with +44.38% gain and $14,438 account balance, DeepSeek Chat V3.1 is in second place with +20.92% gain and $12,092 balance, and Chinese AIs round out the top two. Other North American models such as Gemini 2.5 Pro lost more than 60%. the Chinese model's strengths are in quantitative thinking, risk control and not over-trading, demonstrating the potential of AI in real market decision-making.
DeepAnalyze is an open source agent-based large language model jointly developed by a team from Renmin University of China and Tsinghua University, the first end-to-end solution for autonomous data science. Its core capabilities include automatically completing the whole process of data preparation, analysis, modeling, visualization and report generation, supporting multi-format data sources such as CSV, Excel, etc. without human intervention.DeepAnalyze-8B (8B parameters) outperforms commercial models such as GPT-4o-mini in benchmarks, and with the model weights, code, and training data fully open-sourced, DeepAnalyze can be deployed as a proprietary data science assistant.
Racer launched AI programming product matrix KAT-Coder, covering self-developed models, tools and platforms, supporting more than 20 programming languages and multiple types of development tasks. Its open source version, KAT-Dev-72B-Exp, surpassed GPT and Claude in the SWE-bench list with 74.6%. The model has the ability of code generation, debugging, optimization, etc., is compatible with the mainstream development tools, and shows a strong potential for application in the fields of webpage generation, e-commerce websites, 3D special effects, etc., marking the official entry of KAT into the AI programming track.
DeepSeek-OCR is based on visual language model and "Contextual Optical Compression" technology, which realizes efficient long text processing by converting images into visual tokens. It maintains an accuracy of 97% at 10x compression ratio and 60% at 20x compression ratio. In the test, it surpasses GOT-OCR2.0 with 100 visual tokens and MinerU2.0 with less than 800 tokens, and is able to rebuild the structure of complex documents.
Manus, as a representative of the AI Agent boom in 2025, relies on large models, tool chains and memory technology to realize task execution, but due to the lack of deep cultivation of professional scenarios and closed-loop delivery, it exposes the bubble of "universal Agent". The root of the problem lies in the lack of engineering accumulation and capital-driven short-sightedness, which leads to the stacking of functions but limited intelligence. The industry is turning to vertical fields, such as medical Agent OpenEvidence, which emphasizes deterministic process and data-driven, revealing that the future belongs to the path of "dumb intelligence" that is focused, evaluable, and solid on the ground.
OpenAI releases ChatGPT Atlas, the first AI-native browser, which deeply integrates ChatGPT's intelligent capabilities. Its core features include: real-time AI-assisted web content summarization and interaction, intelligent writing optimization, natural language control of browser operations, personalized memory recommendations, intelligent body mode to automate shopping and booking tasks, and real-time text processing for cursor chat. The browser improves browsing efficiency, automates tasks and reshapes the human-computer interaction experience through AI technology.
Google's Veo3.1 competes with OpenAI's Sora2 in the field of AI video generation; Veo3.1 has the advantage of precise control and high-quality synchronization of audio and video, which is suitable for the creation of professional long videos, while Sora2 has the advantage of smooth and natural dynamic effects and entertainment, which is more suitable for creative short videos. Both have their own advantages, and the choice depends on the specific application scenarios.
Google Gemini 3.0 Pro can generate a fully functional web interface for macOS, Windows, and Linux operating systems in a single HTML file based on simple prompts for a smooth interactive experience. Although its performance in interface replication and functional simulation is excellent, surpassing models such as Claude 4.5 Sonnet, it is still limited to front-end simulation and lacks underlying system logic.
Tencent's hybrid team has launched Hunyuan3D-PolyGen, the industry's first 3D generative large-scale model that meets the standards of art grade, capable of generating professional 3D models that can be used in game development and film and television production, significantly improving the efficiency of artists. The model has significant technological breakthroughs in complex geometry modeling capability and generation stability, supports multiple input methods, and significantly reduces the number of tokens and improves modeling quality through BPT compression technology and reinforcement learning optimization strategy. It can be experienced for free through the Tencent Hybrid 3D platform.
Excel table processing is often vexing due to complex operations, emerging AI tool Shortcut simplifies the process through natural language interaction. It completes complex tasks in 10 minutes in simulated Excel tournaments with an accuracy rate of 80% or more, supporting a wide range of applications from data processing to financial modeling. Natural language input to replace the function syntax, the convenience is significant, but there are still limitations on extremely complex data processing and formatting. Currently in internal testing, Google email users can experience 3 times for free.
MuseSteamer, a multimodal generation model launched by Baidu's commercial R&D team, has achieved the world's first place in VBench's graphic video evaluation, and has made important breakthroughs in the simultaneous generation of Chinese audio and video, refined description system and style control, and has demonstrated superior semantic comprehension capabilities. Despite the lack of lens scheduling ability and slow generation speed, MuseSteamer is still an important milestone in the development of domestic AI video technology, and the Turbo version has been opened for free to experience.
Tencent AI Lab has launched SongGeneration, an open source music generation model, which breaks through the challenges of sound quality, musicality and generation speed through innovative technical architecture and training methods. The model supports four core functions: intelligent text control, precise style following, multi-track generation and timbre cloning, significantly lowering the threshold of music creation. The three-stage training strategy and multi-dimensional human preference alignment further enhance the generation effect. Authoritative evaluation shows that the model ranks first among open source models, close to the level of commercial models, and has been open to experience in Hugging Face and GitHub, helping to popularize intelligent music creation.
AliCloud recently released its latest multimodal AI model, Qwen-VLo, whose image generation and editing capabilities have been highly rated by users, even surpassing GPT-4o. The model has the advantages of enhanced detail capture, single-command image editing, multi-language support, and flexible resolution adaptation, and excels in image recognition, object replacement, and progressive generation. It is now available for free via the Qwen Chat platform.
Embrace your AI assistant and boost your productivity with just one click!
