# OpenBMB releases MiniCPM5-2B, a dense 2B model with open training data, under Apache 2.0

> MiniCPM5-2B is a dense 2B on-device model for coding, tool use and long context, released with its UltraData training sets.

Published 2026-09-06

Canonical: https://shipped.anth.us/open-weights/openbmb-minicpm5-2b/

OpenBMB published MiniCPM5-2B, the second model in its MiniCPM5 series after MiniCPM5-1B. It is a dense 2B transformer, listed at about 2.5 billion parameters on Hugging Face, aimed at on-device and resource-constrained use. Tags on the card include long-context, tool-calling and edge-ai, and the language list is English and Chinese.

The card claims the best results among 2B-class open models in its own comparison set, and says it stays competitive with 4B-class models. Its radar chart averages 53.9 against 51.1 for Qwen3.5-4B. Those are the maker's numbers and its choice of comparison.

OpenBMB also released the data behind it as the UltraData family, including a web pre-training set, code data, 500K agent training samples and RL data. Several checkpoints are listed: final, SFT, mid-train and base. The licence is Apache 2.0.

The card gives no release date; the date used is the Hugging Face repo creation date.

Roughly 2B parameters, with the receipts attached. Training data releases are the rarer half of that sentence.

## Sources
- [Model card on Hugging Face](https://huggingface.co/openbmb/MiniCPM5-2B)
- [Hugging Face repo metadata (creation date, licence, size)](https://huggingface.co/api/models/openbmb/MiniCPM5-2B)

## Editor's first pass (AI-assisted)
- Real 4.5: Apache 2.0, ungated, about 1.3 million downloads; small enough for edge devices.
- Useful 3.5: On-device use with tool calling and long context; the card compares it to 4B models.
- Notable 4: 1,800 likes; ships with datasets and several checkpoints.
- Novel 2.5: Scales up the MiniCPM5-1B recipe; the data release is the extra.