MiniCPM5-2B - Small enough for the device, built to act
MiniCPM5-2B is OpenBMB's dense 2B follow-up to MiniCPM5-1B. It is built for on-device and local use, leads the 2B open-source class, and stays competitive with larger 4B models on coding, math, long context, and agent tasks. Apache 2.0 weights, GGUF, and MLX are up.
Replies
Flowtica Scribe
Hi everyone!
OpenBMB has been steadily betting on open, on-device models for a while now. They just keep shipping solid little models and pushing the same direction forward.
And now, boom — MiniCPM5-2B is the highest-scoring open-weight model under 4B parameters on the latest Artificial Analysis.
At this size, it is already doing useful tool calling, coding, search, and agent work, while still small enough that local deployment is the whole point.
The release includes checkpoints from base training through SFT to the final model, plus 500K agent SFT samples and 80K+ RL samples.
GGUF and MLX builds are already up, and there’s a hosted demo you can try yourself.
2B is getting harder to think of as just a tiny local model!
Veltrix AI
wow wow! Finally! Good one! Actually, are you going to create a separate ready-from-the-box device?