Congratulations to MiniCPM5-2B for topping Hugging Face Trending
🏆 MiniCPM5-2B has officially reached the top of Hugging Face Trending! This result belongs to developers around the world. Thank you for downloading, trying and reporting back so quickly, and for every star and every share.
💡 2B parameters, another step up in overall capability
On the AA Intelligence Index V4.1.1, MiniCPM5-2B ranks first among open-source base models under 4B worldwide, with an average score of 53.9 across 34 benchmarks.
🤖 Stronger on-device agent capability
It supports tool calling, deep search, code generation and other agent tasks, bringing agents further into real devices such as phones, PCs and car cockpits.
🧠 Model, data and RL techniques released together
Beyond the model weights, we are releasing code pre-training data, Agent SFT and RL data, along with the data refinement framework UltraX, the reinforcement learning framework Meshy and the RL strategy JustRL II, to support research, reproduction and application development.
📱 Built for real on-device deployment
Day 0 support is already in place for Intel, Rockchip, Arm and other platforms, along with mainstream inference and fine-tuning frameworks, making local deployment and application development easier.
From 1B to 2B, we keep bringing model capability into more real scenarios.
So local assistants, agents and tool calls run in everyday devices and workflows.
MiniCPM5-2B is available for download. Whether it is a deployment experience, an agent application, a real-world case or a log of pitfalls, share it with us in the community group. Let's see what a 2B model can do next! ❤️
🔗 HF:
https://huggingface.co/openbmb/MiniCPM5-2B
🔗 GitHub:
https://github.com/OpenBMB/MiniCPM
🔗 Meshy:
https://github.com/OpenBMB/Meshy
🔗 ModelScope:
https://modelscope.cn/models/OpenBMB/MiniCPM5-2B
🔗 JustRL Ⅱ:
https://panhaoxuan.notion.site/justrl-ii-small-llms-to-128k-reasoning-with-a-critic-cn