openbmb launches MiniCPM5-2B, a 2B on-device model with Apache-2.0 weights
openbmb released MiniCPM5-2B, a dense 2B Transformer for on-device and resource-constrained deployment, on Hugging Face under Apache-2.0. The model is positioned as 2B-class open-source SOTA and competitive with 4B-class models in coding, math, long-context, tool use, and agentic tasks.
China context
- Outside China
- Open weights · huggingface.co
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Builders outside China can download the Apache-2.0 licensed weights from Hugging Face and integrate the model into local or edge deployments without API costs.
- For investors
- Investors can track whether MiniCPM5-2B's claimed SOTA performance leads to adoption in on-device AI products, which could pressure larger model providers in the small-model segment.
openbmb released MiniCPM5-2B, the second model in the MiniCPM5 series, on Hugging Face. It is a dense 2B Transformer designed for on-device, local deployment, and resource-constrained scenarios. The model is described as achieving 2B-class open-source SOTA and remaining competitive with 4B-class models overall, with advantages in coding, mathematics, long-context understanding, tool use, and agentic tasks. The release includes links to a tech report, Chinese wiki, GitHub repository, UltraData, and an online demo.
MiniCPM5-2B is a dense 2B Transformer that scales up the training recipe from MiniCPM5-1B. The model card includes a capability radar chart comparing MiniCPM5-2B with 4B-class models across dimensions such as code reasoning, math reasoning, instruction following, general knowledge, long context, tool use, coding agent, search agent, and general agent. The radar chart is normalized independently to 100 percent per axis, so absolute performance levels are not directly comparable from the chart alone.
Developers outside China gain an Apache-2.0 licensed 2B model that can be deployed locally, reducing reliance on larger cloud-hosted models for on-device applications. Competitors in the small-model segment now face a new benchmark for 2B-class performance in coding, math, and agentic tasks.
The Apache-2.0 license allows commercial use, and the model's focus on on-device and resource-constrained deployment could lower infrastructure costs for developers targeting edge devices. The claimed competitive performance with 4B-class models may reduce the need for larger, more expensive models in certain applications.
The next verifiable signal is whether independent benchmarks confirm the claimed SOTA performance and whether the model sees adoption in on-device applications. The release of the tech report and GitHub repository may provide further details on training data and methodology.