Baidu AI
104 posts
@BaiduAI_News
Discover how Baidu is bringing AI to life.
WAIC 2026 is almost here, and we're bringing quite the AI lineup. 👀 From our complete lineup of AI agents to our full-stack AI capabilities, and upgraded AI app experiences, there's plenty to explore. Come see us July 17–20 at the Shanghai World Expo Exhibition and Convention Center, Hall H1, Booth B107.
We are happy to host a group of local reporters for an Apollo Go ride on Airport Island today, alongside the Hong Kong Transport Department. What initially began on Airport Island has since expanded to North Lantau, Kowloon East, and Hong Kong Island South. Testing has gone smoothly throughout as Apollo Go adapted to the city's right-hand-drive, left-hand-traffic layout and held steady across changing weather, busy traffic, and varying network conditions. By the end of May, that adds up to 240K+ km of safe driving. The technology powering Apollo Go is ready for what comes next. 🛣️
AI for Good in action: Three of our AI use cases were recognized as winning cases at the 2026 @AIforGood Global Summit, and featured in the latest Innovate for Impact Report. 🌍✨ > Apollo Go: Scaling Sustainable and Inclusive Urban Mobility > Miaoda & MeDo: Democratizing Software Creation through AI-Powered No-Code > PaddlePaddle: AI-Powered Restoration of Thangka Sacred Art Different fields, same idea: AI should help more people create, preserve, and move through the world. See how each one is making an impact ↓
Real roads are full of moving parts, but Apollo Go anticipates how situations may unfold and responds smoothly in real time. Watch Apollo Go navigate the unexpected.
WorldClaw is partnering with Baidu AI Cloud @Baidu_Inc to bring the ERNIE 5.0 series to WorldRouter: one API connecting developers to Baidu’s full model stack, alongside 300+ models already on worldclaw platform. This opens frontier-scale capability to the fast-growing agent ecosystem of startups, model builders, enterprises and independent developers. To mark the launch, all AI models are 30% off (USD / 1M tokens), including: ERNIE 5.0 - $0.98 in / $3.92 out Kimi-K2.6 - $0.67 in / $2.80 out DeepSeek-V4-Flash - $0.10 in / $0.20 out DeepSeek-V4-Pro - $1.22 in / $2.44 out GLM-5.2 - $0.98 in / $3.08 out Seedance: 5% off. router.worldclaw.ai
Unlimited-OCR has surpassed 1M downloads on Hugging Face 🚀 ❤️ Huge thanks to everyone supporting the project 👀 Next up: multilingual and 128k context versions GitHub: github.com/baidu/Unlimite… HF: huggingface.co/baidu/Unlimite…
Excited to share our #ICML2026 paper: AdaGC! Loss spikes plague LLM pretraining. AdaGC uses a tensor-wise adaptive EMA mechanism to completely eliminate recurrent spikes on ERNIE 10B-A1.4B. 🚀Zero spike score! 📈Up to +2.48% accuracy ⚡Saves 4.48% GPU hours #LLM #NLP #AI
Tokyo builders, we’re coming to you. On July 17, MeDo is teaming up with @She_Nicest for a hands-on Builder Meetup in Tokyo. Bring a rough idea, a daily problem, or an app you’ve always wanted to try building. With MeDo, you’ll turn it into a working demo through natural language. What to expect: ✨AI creator case sharing ✨MeDo tool demo ✨60-minute hands-on building ✨Demo sharing & feedback 📍 Le Wagon Meguro Campus, Tokyo 🗓 July 17, 2026 🕢 7:30pm – 9:30pm JST 🎟 Register on Luma (luma.com/rkhe7vcs) Build your first idea with MeDo. 👉 Try MeDo: medo.dev #MeDo #SheNicest #TokyoMeetup #VibeCoding #AIBuilder #NoCode #Hackathon #BuildWithMeDo
🎉 Unlimited-OCR from @Baidu_Inc now runs in vLLM. One-shot parsing of entire books with constant KV cache, powered by Reference Sliding Window Attention (R-SWA). 🧠 R-SWA keeps KV cache fixed throughout decoding — no memory blowup, no slowdown, no matter how long the output gets. 📄 Transcribe 40+ pages in a single forward pass under a 32K context budget, with remarkably low edit distance even at scale. 🪶 35% faster than DeepSeek-OCR at 6K output tokens, with fully constant TPS and GPU memory. 🔗 Recipe: recipes.vllm.ai/baidu/Unlimite… 🤗 Weights: huggingface.co/baidu/Unlimite… 🙏 Thanks to the @BaiduAI_News team for the collaboration.
🌍 Build with MeDo: Global Hackathon Recap 🏆 A quick look back at our Global Hackathon. A huge thank you to our incredible community across 160+ countries for building with us. Here is the final impact we made together: 🔥 The Final Numbers: 🌐 10M+ Total Reach 👥 8,000+ Participants 📦 2,262 Submissions 🗺️ 160+ Countries 📊 Participants by Track: 💼 26.8% — Work & Productivity 🪄 20.7% — Surprise Us! 🎮 20.1% — Lifestyle & Game 🎓 18.6% — Learning & Education 🛒 13.9% — Business & E-commerce From everyday utilities to fully functional business tools, these submissions proved that creators can bring their ideas to life natively with MeDo. #MeDo #AI #VibeCoding #NoCode #Hackathon #BuildWithMeDo
Real5-OmniDocBench is accepted to ECCV 2026! 🎉 We physically reconstructed OmniDocBench v1.5 across 5 real-world scenarios. Key insight: Real robustness ≠ big params, it's domain-smart design. Paper: arxiv.org/abs/2603.04205 Dataset: huggingface.co/datasets/Paddl…
