OpenBMB's 2B model punches well above its weight
OpenBMB just open-sourced MiniCPM5-2B, a compact 2-billion-parameter dense model that punches above 4B-class models like Qwen3.5-4B across coding and math benchmarks. Weighing only around 5GB, it easily runs locally on consumer edge hardware, and the team also released the full high-quality dataset behind it.
MiniCPM5-2B is OpenBMB's new 2-billion-parameter dense language model designed specifically for local inference on consumer edge devices and laptops.
Packing exactly 2,516,756,480 parameters into a roughly 5GB footprint, it easily fits inside everyday hardware without needing costly GPU setups.
What it can do
Despite its compact 2B size, MiniCPM5-2B punches well above its weight class. In internal and standardized tests, it goes head-to-head with and often outperforms 4B-class rivals like Qwen3.5-4B and granite-4.2-3B across coding, math reasoning, instruction following, and broad knowledge tasks.
It ranked #1 among open-source models under 4B parameters on the Artificial Analysis Intelligence Index with a score of 23, alongside an Agentic Index score of 20. It also supports an expansive context window of up to 131,072 tokens, and handles multilingual queries like Japanese with clear sentence structure rather than degraded output.
How it was built and what is open
The secret behind its density is training data, and OpenBMB did not hold it back. Along with the model checkpoints, the team open-sourced its full UltraData family.
That release includes the UltraX web pre-training data, tiered UltraData-Code, 500,000 agent samples in UltraData-SFT-Agent-2609, and over 80,000 reasoning samples in UltraData-RL-2609. The weights are published under the Apache 2.0 license across Hugging Face and ModelScope, featuring BF16 base versions, GGUF builds for llama.cpp, 4-bit MLX files for Apple Silicon, and ready-to-run vLLM and SGLang setups. You can also try it immediately via a Hugging Face Space web demo.
Why it matters
If you build local tools or run offline agents, MiniCPM5-2B is definitely worth spinning up. A 2B model that can genuinely reason through code and long context without instantly hallucinating makes smart on-device copilots far more realistic on ordinary laptops and phones.
Releasing the training datasets alongside the weights is also a major win for the community. Instead of just another closed edge checkpoint, builders now have both the model and the high-grade recipe needed to tune custom lightweight agents.
Sources (3)
- huggingface.co(opens in a new tab)Originalhuggingface.co
- gigazine.net(opens in a new tab)Reportgigazine.net
- youtube.com(opens in a new tab)Reportyoutube.com