ENLatest AI NewsAliyun BaiLian Launches HappyOyster 1.0: Open-World Model Supporting Real-Time Interaction Begins Gray TestingAliyun BaiLian launches the open-world model HappyOyster 1.0, which supports real-time construction, exploration, and interactive generation of AI digital worlds. The model learns state transition rules, differing from traditional unidirectional generation. It has released SDKs for Android, iOS, and Web platforms and has started gray-scale testing.just now5.1KAlibaba Launches Token Plan Personal Version, Qwen3.8-Max-Preview Goes Live SimultaneouslyAlibaba officially launches the Token Plan personal version and opens up early access to the 2.4 trillion parameter large model Qwen3.8-Max-Preview. This model performs outstandingly in coding and office scenarios. The official version will be open-sourced. Users can obtain model usage rights through Token Plan. Alibaba also launches a limited-time promotion: Lite plan for 39 yuan/month, Standard plan for 139 yuan/month, Pro plan for 499 yuan/month; Team versions offer standard and advanced options.just now3.9KGoogle Upgrades Gemini Enterprise: Optimizing Interface and Advancing Workspace Connector RedesignGoogle is advancing the enterprise AI assistant Gemini Enterprise update, adopting a vibrant visual design consistent with the consumer version, unifying the enterprise and personal user experience, highlighting its strategy to align enterprise AI with the billions of users in the Gemini ecosystem. Additionally, the Google Workspace connector has been upgraded, and some users will need to reauthorize to continue using it.just now9.9KTencent Hunyuan Hy3 Free Offer Extended to August 5th, WorkBuddy and CodeBuddy Users Can Continue to Use for FreeTencent’s Hunyuan Hy3 model extends its free trial to Aug 5 due to positive user feedback, targeting WorkBuddy and CodeBuddy users to test intelligent code generation and office assistance capabilities…..just now6.4KAfter BrowseComp was brushed to 90%, Meituan LongCat launched LoHoSearch: Frontline models collectively dropped back to less than 30%“Search Agent Evaluation Benchmark BrowseComp was quickly overwhelmed,” its performance soared from 30% to 90% and gradually became ineffective. On July 17, Meituan LongCat released a new benchmark LoHoSearch, which generates difficult problems based on a Wikipedia knowledge graph containing 7.62 million entities, aiming to push the evaluation back into a high-difficulty area and reset the standard for search agent capabilities.just now9.2KYang Zhilin at GTC2026 Reveals: Moon’s Shadow Has Replaced Adam, Full Attention, and Residual Connections, All Open SourceMoonshot AI CEO Yang Zhilin says making open-source models match closed-source ones requires rebuilding Transformer foundations, not just scaling parameters and compute. They redesigned the Adam optimizer, attention, and residual connections, reshaping LLM training…..just now8.6KNot enough with whisper.cpp and ONNX? The author of Handy opens up transcribe.cpp: 60+ transcription models, GPU acceleration, plug-and-play alternativeLocal speech transcription developers struggle with maintaining multiple engines like whisper.cpp, ONNX, and MLX. On July 19, Handy creator sebjones released transcribe.cpp v0.1.0, a ggml-based unified cross-platform solution supporting all latest models, directly addressing engine fragmentation…..just now8.1KKunlun Tech has designated 2026 as the Year of the World Model: Matrix-Game 3.5 can generate a 5B model in real-time on a single card, achieving 20FPSKunlun Tech announced at WAIC that 2026 will be the inaugural year of world models, unveiling Matrix-Game3.5 and two other products. The model features patch-level memory injection, enabling recall of image fragments for real-time interaction…..just now9.3KEpoch AI Tests Three AI Text Detectors: Nearly 30% of Content Missed After Mimicking Human Writing StylesEpoch AI research shows that mainstream AI text detectors can nearly perfectly identify ordinary AI-generated content, but when large language models deliberately mimic the writing styles of specific authors, their accuracy drops significantly, with scientific writing being the hardest to detect. The experiment tested three tools: Pangram, GPTZero, and Originality.ai. It used 495 human-written texts covering blogs, novels, and scientific writing (all created before ChatGPT existed), and found that style imitation can effectively evade detection.just now7.8KGoogle DeepMind Launches GenCeption: Achieving Multiple Computer Vision Breakthroughs with a Video-Based Generation ModelGoogle DeepMind released GenCeption, a pretrained video generation model that unifies depth estimation, image segmentation, 3D pose estimation, etc. Its performance rivals or surpasses specialized models with much less training data. This challenges the landscape dominated by models like Segment Anything and Depth Anything…..just now4.9KMingxi Intelligent Open Sources MiniCPM-Robot: Full Rel