AI News

ByteDance's Seed foundation model unit reorganized: Seed Foundation Model created four first-level departments reporting to Wu Yonghui. It consolidates pretraining data, RL and post-training resources, separating B-end and C-end product model capabilities. Pretrain Data led by Li Chenggang integrates text, coding, vision, speech data teams to support new Omni model.....

The DeepSeek toolchain is upgraded to version 0.1.1-rc.1, adding the visual model experience version V4Flash Vision-Exp on top of V4Flash and V4Pro. Previously, version 0.1.0-rc.8 had already enabled multimodal capabilities, supporting native image requests and mixed text-image input with commands such as /goal and /plan.

TaoTie Innovation launches the new iO series. The first TaoTie iO is positioned as an 'AI-enhanced glasses for humans', priced at 2499 yuan, with a launch price starting at 1996 yuan. It supports prescription lenses from 0 to 1000 degrees and is now available on Jingdong, Tmall, Douyin, and offline channels. In the first quarter of 2026, TaoTie ranked first in AR glasses shipments in China and globally. The new product features dual-eye display and Blue Lake waveguide technology, with a 93% optical clarity (OC area) and over 98% in non-display areas.

iFLYTEK revealed at its 2026 Semi-Annual Performance Briefing that it will officially launch a new main general-purpose large model based on domestically-produced computing power during this year's 1024 Developer Festival; a phased version will be released in late August for early market and developer experience, with an established technical implementation schedule.

Kakao announces its restructuring: a new company called 'Kakao AI' will focus on KakaoTalk and artificial intelligence. The original company will be renamed 'Kakao X' to handle investments, fintech, content, and mobility. The reorganization aims to operate independently, accelerate AI implementation, and improve capital allocation efficiency. The net assets will be split in a 36% and 64% ratio.

OpenAI opens Codex Harness core capabilities to developers. It powers Codex agents, connecting models, users and tools, and managing tasks, context, sessions and code operations. Codex CLI and core code were open-sourced, with ongoing improvements to Harness and App Server. It is not just code generation but a complete agent execution framework.....

Alibaba launches the GUI Intelligent Agent Foundation Model Qwen-UI-Agent, centered on the real world, covering mobile, desktop, web, and deep search, aiming to break through the limitations of simulated testing and achieve seamless operation on complex real devices. It achieved a high score of 82.1% in the MobileWorld mobile test, leading multiple mainstream models, with comprehensive performance comprehensively surpassing industry benchmarks.

When enterprises build AI, core confidential data is often buried in long documents, financial reports, legal contracts, and other files. Traditional one-time RAG only retrieves fixed text blocks, lacking iterative reasoning, making it difficult to handle complex long documents and cross-source comparisons. To address this, Mistral AI launched the Agentic upgrade solution, enhancing iterative reasoning capabilities and solving the challenge of information extraction from long documents.

Liquid AI and Hugging Face released three LFM2.5 DSpark draft checkpoints (1.2B, 2.6B, 8B-A1B). A new speculative decoding path boosts inference throughput up to 3.18x on GPU and accelerates edge devices, without altering output quality.....

SenseTime opensources the 8B lightweight unified multimodal model SenseNova U1.5Lite, which demonstrates superior performance in core visual tasks, comparable to large-scale commercial models in handling complex layouts and text rendering; it natively supports 3-4K context length, enabling processing of multiple constraints such as subject, quantity, spatial relationships, text, layout, and style simultaneously.

SAP and Team Liquid jointly developed an AI voice application that can automatically process multi-channel internal voice communications during matches and training, accurately separate individual player voice streams, and complete transcription, speaker identification, and emotion analysis. It transforms scattered conversations into structured match data, helping teams analyze matches, optimize strategies, and improve player performance.

Doubao's Work Task Mode has been upgraded with three new features: Skills, Connectors, and Work Partners, enhancing collaborative work capabilities. The platform has already launched over 200 Skills and Connectors, supporting users in building custom Skills and standardized workflows. 'Skills' can consolidate personalized work methods, such as saving a weekly report format once and reusing it every week; 'Connectors' integrate Doubao with various office systems, improving the standardization and efficiency of work processes.

xAI announced that the application building feature of Grok Build has completed testing and is now officially available to all subscription users, covering web, iOS, and Android. Users can conveniently create applications or games by simply describing their needs in natural language when interacting with Grok, significantly lowering the barrier to AI native application development and promoting the deep integration of AI applications into productivity.

The Alibaba Cloud Qwen AI Platform has recently updated its model matrix, officially integrating the GLM-5.3 from Zhipu and the DeepSeek-V4-Pro, with APIs now available. Both models focus on programming and intelligent agent scenarios: GLM-5.3 enhances programming, long-term tasks, and cybersecurity capabilities; DeepSeek-V4-Pro improves intelligent agent interaction, suitable for code development and multi-step complex task processing.

Anthropic announced that the computer operation, Skills API, and Files API of the Claude platform are fully open, with added browser usage tools and optimized underlying capabilities, helping enterprises build operational software, apply team expertise, and output production-level agents.

Pew Research Center reported Aug 20 that over one-third of webpages published after ChatGPT's launch show signs of AI writing or editing, indicating generative AI is rapidly reshaping online content. The study analyzed about 500,000 English pages from Common Crawl over five years using Open Pangram; about 10% of 10,000 randomly sampled pages had clear AI traces, partly because the sample includes many pre-AI pages.....

Waymo has deeply integrated Google's Gemini AI assistant into its custom autonomous taxi Ojai, driving the intelligent upgrade of driverless travel. The model is now fully available to all regular users in Los Angeles, Phoenix, and San Francisco. The company also plans to further expand its deployment this year.

Adobe launches Firefly audio tools for music, voice, and sound effects, enabling commercially safe creation on desktop and mobile within video workflows. A survey shows all creators use music in videos; 32.7% already use AI-generated music. Main barriers: 43.2% cite legal copyright risks, 38.9% licensing costs.....

On August 20, the Lhasa branch of the 2026 Global Digital Economy Conference held the 'Digital Najin · Intelligence Summit on the Plateau' forum. The Lhasa Njajin Digital Intelligence Research Institute was officially established, and the 'Yangguang Qingyan' general Tibetan large model was officially released. The model first appeared on November 19, 2025, and has been registered with the Cyberspace Administration of China. It was officially launched after nearly two months.

Rising demand for AI training data has driven rapid growth for annotation startups. Micro1, founded four years ago, increased its annualized revenue from $1 billion to $5 billion in the past 8 months, with net income of approximately $1.5 to $2 billion. Although its scale is still smaller than Mercor and Handshake (annualized total revenue of $2 billion and $1 billion respectively), the growth highlights that the AI training data market is forming a sufficiently large scale.