🔥 Spotlight
Mysterious Model Ox Alpha Unexpectedly Launches on OpenRouter and OpenCode : OpenRouter and OpenCode unexpectedly launched a mysterious stealth model codenamed “Ox Alpha”, offering a 1 million token context window and native support for text, image, and video inputs. Designed for long-horizon Agent coding, complex reasoning, and real-world production environments, the model offers a free testing quota of 100 trillion tokens daily. Multiple developer benchmarks show its coding and reasoning capabilities surpass Fable 5 and GPT-5.6 Sol. Wild community speculation regarding its true identity is ongoing, with main guesses pointing to Zhipu GLM-5.4/5.5 or Xiaomi MiMo V3 (Source: OpenRouter, 36Kr)
DeepSeek Officially Launches Multimodal Vision Model V4-Flash-Vision-Exp : DeepSeek’s official API platform unexpectedly released its first native vision understanding model, DeepSeek-V4-Flash-Vision-Exp. While maintaining V4-Flash’s strong text capabilities, its multimodal Agent performance has improved significantly, approaching Opus 4.8. Images are converted into tokens based on dimensions (capped at 384 tokens per image). Pricing is completely identical to V4-Flash (equivalent to around 1 RMB per 1,000 images), with concurrent free Files API access and Harness 0.1.1 integration, significantly lowering the development and operational costs of vision Agents (Source: DeepSeek, Synced)
Google Introduces Gemini 3.7 Flash, Refreshing Multiple Agent Benchmarks : Google officially introduced Gemini 3.7 Flash, featuring a 50% price reduction compared to 3.6 Flash alongside significant algorithmic improvements in intelligence. In Artificial Analysis’s AA-AnalystAgent real-world data analysis benchmark, Gemini 3.7 Flash took top spot, completing tasks 60% to 90% faster than competitors. Meanwhile, it scored 84.6% on the ARC-AGI-2 evaluation at minimal cost, demonstrating extreme cost-effectiveness and visual Agent capabilities (Source: demishassabis, Google Research Blog)
OpenAI Establishes Foundation and Launches AI Futures Initiative : OpenAI officially established the OpenAI Foundation and launched the AI Futures project, aiming to convert AGI capabilities into broad societal public good. The foundation focuses on life sciences (such as protein design and drug discovery) and AI societal resilience (defending against biological/cyber threats). It has deployed hundreds of millions of dollars in its first five months, with plans to commit billions in the future, including a $17.2 million donation to SecureBio to build an early warning system. This move marks top AI labs accelerating the construction of public infrastructure to prevent tech misuse and societal risks (Source: OpenAI News, woj_zaremba)
🎯 Developments
NVIDIA Introduces AVO Agent, Scoring a Perfect 100% on ARC-AGI-3 : NVIDIA AI team released AVO, a general-purpose programming and interactive reasoning agent. Without any explicit instructions, rules, or preset goals, the AVO Harness powered by Claude Opus 5 completed all 183 levels across 25 environments in the ARC-AGI-3 interactive reasoning benchmark, achieving a perfect score of 100% while improving action execution efficiency by 12% over similar systems (Source: ClementDelangue)
Alibaba Open-Sources Qwen-UI-Agent Foundation Model : Alibaba released Qwen-UI-Agent (available in 27B / 35B-A3B / 4B versions), a GUI agent foundation model for mobile, desktop, and web execution, trained on over 100 real phone environments. The model unifies GUI clicks and CLI action spaces, outperforming GPT-5.6 Sol and Opus 4.8 across 5 core GUI benchmarks (Source: 36Kr)
Xiaohongshu Open-Sources FireRedTTS3 Unified Speech Generation and Editing Model : Xiaohongshu released its next-generation speech model FireRedTTS3. Based on its pioneering RedAE semantically enhanced continuous representation, it unifies zero-shot cloning, natural language voice design, and precise local speech editing across 24 languages and 21 Chinese dialects within a single model, sweeping top spots across four public benchmarks including Seed-TTS-Eval (Source: Synced)
Replit Launches Free Mode Powered by GPT-5.6 Luna : Replit partnered with OpenAI to launch Free Mode powered by GPT-5.6 Luna, allowing users worldwide to experience interactive AI programming and Agent app development for free, significantly lowering the barrier to building complete software from scratch (Source: amasad)
Meta Releases Muse Spark 1.2 Native Omni-Modal Model : Meta officially released Muse Spark 1.2, featuring strong visual code generation, robotics planning, and audio-video understanding capabilities. In Design Arena evaluations, it won top spot on the Video-to-Website leaderboard and ranked near the top on Image-to-HTML, demonstrating high omni-modal practical value (Source: alexandr_wang)
OpenAI Previews Native Transparent Background Generation in GPT-Image-2 : OpenAI introduced a preview of the background=transparent parameter in the GPT-Image-2 API, enabling the model to directly output transparent PNG layers with alpha channels during generation. Its results on complex textures like glass transparency and fine edge filaments far surpass traditional post-processing cutout techniques (Source: THE DECODER)
Adobe Firefly Integrates Google Gemini Omni Flash and Launches AI Audio Tools : Adobe announced the integration of Google’s Gemini Omni Flash multimodal video model into its Firefly creative platform, while fully launching three commercially safe AI audio generation tools: Generate Music, Generate Speech, and Generate Sound Effects (Source: THE DECODER)
Google Releases TIPS Spatially-Aware Vision-Language Model : Google open-sourced the spatially-aware vision-language model TIPS (and TIPSv2) on Hugging Face, specifically designed for dense vision understanding tasks such as image segmentation and depth estimation (Source: gabriberton)
Runway Introduces 16-bit HDR Video Enhancement Model Ruby : Runway released the video model Ruby, supporting the conversion of SDR video to 16-bit HDR ProRes/EXR sequences, improving color depth and image quality for both generated videos and uploaded assets (Source: c_valenzuelab)
NetEase Netbee Explores AI Interactive Community and Next-Gen UGC Supply : NetEase Netbee is deeply integrating AI capabilities based on open-source and proprietary models into youth interest communities, launching the “Wannao Lobster” personal assistant and AI interactive games/interactive graphic content, driving community content formats toward playable and re-creatable interactive UGC (Source: 36Kr)
🧰 Tools
MiniMax Releases MiniMax Design Video Generation and Creation Platform : MiniMax launched MiniMax Design, a platform targeting commercial video production. With Agents as its core entry point, it integrates scripting, storyboarding, generation, dubbing, and editing onto a single canvas, introducing reusable Skill mechanisms to directly benchmark against commercial video pipelines of Adobe and Canva (Source: 36Kr)
DP Technology Releases “Bohr Science Space” Desktop AI Research Collaboration Environment : DP Technology launched the desktop application “Bohr Science Space,” featuring built-in domain AI experts such as SciMaster, BioMaster, and PharmMaster. It connects the entire workflow of literature research, hypothesis deduction, data computation, and paper writing, helping researchers take over tedious research labor (Source: QbitAI)
Lindy Launches Chrome Extension for Intelligent Inbox Management : AI Agent platform Lindy launched a Chrome extension embedded directly into Gmail, enabling automatic nightly email summaries, reply drafting, and automatic label classification using personal memory bases (Source: omarsar0)
ChatGPT / Codex Integrates ExaAI Search Plugin : OpenAI integrated the ExaAI plugin into ChatGPT Work and Codex, enabling AI Agents to directly search over 100 billion web pages, academic papers, and enterprise documents across the web (Source: OpenAIDevs)
OpenHistory Open-Sources Native Mac Workflow Tracking App : Developers open-sourced OpenHistory, an application that uses on-device Mac models to automatically record work trails and generate summaries, supporting collaboration with local Agents via secure MCP protocols (Source: zachtratar)
OpenBot Open-Sources Cross-Harness Grok Bot Alternative : The CopilotKit team open-sourced OpenBot, providing a Grok Bot alternative compatible with any Agent Harness, supporting generative UI, computer control, Agent-Human collaboration, and private data retention (Source: hwchase17)
📚 Research & Learning
Pleias Researchers Propose 0.63M Parameter Ultra-Tiny Verifier : Researchers explored the lower parameter bound for verifier models. Requiring only 2 minutes of pre-training on a single H100 GPU, a 630k-parameter ultra-tiny model achieved performance on par with 7B models on specific verification tasks (Source: Dorialexander)
Spark-to-Paper End-to-End Open-Source Paper Generation System : A research team introduced Spark-to-Paper, an open-source paper generation system based on programming assistants. Containing 13 composable skills, it can autonomously run experiments, plot vector figures, verify data, and output compilable LaTeX papers directly from research ideas, achieving a 92% detection rate for false conclusions (Source: Synced)
ZJU and Baidu Propose BEACON Reinforcement Learning Framework for Long-Horizon Agents (ICML 2026) : Zhejiang University and Baidu proposed BEACON, a milestone-guided policy learning framework. By segmenting trajectories at milestone boundaries and combining dual-scale advantage estimation, it resolves credit misassignment in long-horizon Agent tasks, boosting the success rate on ALFWorld long-horizon tasks from 53.5% (GRPO) to 92.9% (Source: Synced)
AutoResearch Evaluation Report: Lack of “Metacognitive Closed-Loop” in Agents Causes Research Failure : Stanford, Prentis AI, and other institutions diagnosed 800 Agent research trajectories, finding that 92.1% of failures stemmed from non-engineering cognitive and logical errors. 82.5% of trajectories showed a metacognitive gap of “noticing issues but failing to correct them,” revealing the bottleneck in moving automated research from simple engineering execution to deep decision-making (Source: Synced)
Patronus AI Open-Sources FigmaTrace Design Computer-Use Dataset : Patronus AI open-sourced FigmaTrace, the first computer-use dataset aimed at UI/UX design, capturing over 200 hours and 3,400+ complex operation trajectories of real designers in Figma (Source: rebeccatqian)
LangChain Founder Releases Managed Deep Agents Tutorial Series : Harrison Chase released a video and tutorial series on Managed Deep Agents, systematically explaining how to build, deploy, and manage deep Agent Harnesses and execution environments in production (Source: hwchase17)
Google Proposes Pandora’s Router Smart Model Routing Theory : The Google DeepMind team formalized multi-model routing as a “Pandora’s Box” search problem. By balancing compute evaluation costs and expert gains, it drastically reduces calls to expensive estimators while matching full search quality (Source: dair_ai)
💼 Business
GPT-5.6 Sol Drives 82% QoQ Surge in OpenAI Enterprise Spending for Q3 : According to latest Ramp data and analysis, driven by the strong performance of GPT-5.6 Sol, OpenAI’s enterprise API spending QoQ growth reached 82% in Q3 2026 to date, surpassing Anthropic’s 76% and boosting Q3 revenue by 35% (Source: THE DECODER, GavinSBaker)
River AI Secures $110M / $1.1B Funding Round with Participation from Cisco Investments : Cisco Investments announced strategic participation in decentralized personal AI startup River AI’s $1.1B funding round, jointly advancing user-owned AI infrastructure (Source: ibab)
Hark Partners with AT&T to Launch AI-Native Consumer Hardware : AI hardware startup Hark announced a strategic partnership with US telecom giant AT&T. AT&T invested in Hark while providing network and device certification support, as both parties co-launch next-generation AI-native personal smart devices (Source: adcock_brett)
🌟 Community
Pangram Analysis: RLHF Post-Training Guardrails Cause “Mode Collapse”, Making AI Text Easier to Detect : AI detection firm Pangram noted that un-finetuned base LLMs originally possessed language diversity comparable to humans. However, behavioral rules imposed by RLHF and safety guardrails lead to Mode Collapse, causing models to favor fixed phrasing, which ironically becomes the primary characteristic enabling high-precision detection of AI-generated text (Source: THE DECODER)
💡 Miscellaneous
JavaScript Runtime Bun 1.4 Stable Released (Completely Rewritten in Rust) : Following its acquisition by Anthropic, Bun officially released the stable version of Bun 1.4, completely rewritten in Rust. It fixes over 2,900 issues, eliminates memory leaks, and features built-in image processing, browser automation, and HTTP/3 deployment, with monthly downloads exceeding 20 million (Source: 36Kr)
RayNeo Releases Lightweight RayNeo iO AI Glasses : RayNeo released the RayNeo iO AI glasses, redesigning optics and frame forms to reduce weight to 34g, supporting two-day battery life and prescription lenses. Powered by an all-day memory engine and multi-model backends such as DeepSeek V4 Pro / Qwen 3.7 Max, it provides proactive knowledge, memory, and real-time translation enhancements (Source: QbitAI)
US-China AI Competition Intensifies Amid Pax Silica Alliance Boundaries : US draft documents require allies to explicitly choose sides in the US-China AI race. Nations joining China’s AI Cooperation Organization (WAICO) will be excluded from the Pax Silica alliance, highlighting the political risks of fragmentation in global AI infrastructure and supply chains (Source: THE DECODER)