In China this week, three real tests: → Pharmacy package: full ingredient and dosage breakdown in English → Restaurant menu: every dish translated with cuisine notes → Metro sign: line, transfer, and…

And the benchmark numbers are actually solid. Compared to Qwen3.5-0.8B, MiniCPM-V4.6 scores higher on: • OpenCompass • OCRBench • RefCOCO • HallusionBench • MUIRBench While maintaining competitive…

在 Github 上发现了一个由清华 OpenBMB 出的口袋级多模态模型， MiniCPM-V，1.3B 参数，能直接跑在手机上。 GitHub链接： 主要功能： 1. 支持单图、多图、视频理解，不只是图片描述 2. intra-ViT 早期压缩，视觉编码计算成本降低 50%+ 3. 可部署于 iOS、Android、HarmonyOS，移动端有现成 App 4.支持…

🚀MiniCPM-V 4.6 hits #1 on @huggingface Trending! 🏆 Huge thanks to the community for the incredible support!😘😘 🔗👇Try it: 🤗 Hugging Face: 💻 GitHub: Modelscope: Web Demo: App Demo: What makes…

🤗 Hugging Face: 💻 GitHub: Modelscope: 🌐 Web Demo: App Demo:

This is what the shift looks like. Not bigger models with bigger bills. Smaller models with better architecture, running anywhere. MiniCPM-V 4.6 beats Qwen3.5-0.8B across benchmarks at 1/19th the…

MiniCPM-V4.6 1.3B is now open source. Beats Qwen3.5-0.8B on OCRBench, RefCOCO, HallusionBench and MUIRBench: - ~50% less visual encoding compute - 4×/16× hybrid compression modes - Runs on llama.cpp,…

6/ This isn’t just a phone demo either. At industrial scale, MiniCPM-V already handles high-concurrency production workloads. Tasks like OCR, captioning, tagging, and ASR — processed in real time, at…