開源模型／端側 AI
MiniCPM5-2B Opens Training Data and On-Device Deployment Formats, but Quantized-Version Sampling Defaults Draw Scrutiny
OpenBMB has released a 2B dense model with a 131K-token context window and tool-calling capabilities, along with staged checkpoints, training data, and multiple quantization formats. Third-party evaluations affirm its competitiveness within its class, but community testing suggests the official sampling settings may cause GGUF quantizations to fall into repetitive generation loops.
OpenBMB has released [MiniCPM5-2B](https://huggingface.co/openbmb/MiniCPM5-2B), positioning it as a text reasoning and agent model designed to run on-device. It uses the standard `LlamaForCausalLM` architecture, with 42 layers and approximately 2.52 billion parameters, including about 1.98 billion non-embedding parameters. Its GQA configuration has 16 query heads and 2 KV heads, and it supports a maximum context length of 131,072 tokens. In addition to Apache 2.0-licensed weights, the team provides Base, Midtrain, SFT, and final RL/OPD checkpoints, as well as GGUF, MLX, GPTQ, LiteRT, and DSpark draft models.
What makes this release particularly valuable for research is that both the data and the post-training pipeline have been opened. The official materials list the UltraX web pretraining dataset, tiered code data, 500,000 agent SFT samples, and more than 80,000 RL samples covering mathematics, programming, and long-context tasks. Post-training begins with SFT, followed by training specialist models for mathematics, programming, agents, and other domains. Finally, On-Policy Distillation merges 16 RL experts. The method derives an advantage signal using reverse KL divergence over the full token distributions of the teacher and student models, rather than relying solely on scores from answer verifiers.
The model's average score across the benchmarks selected by the developers is 53.9, higher than several larger comparison models in the published table. The independent [Artificial Analysis](https://artificialanalysis.ai/models/minicpm5-2b/) page currently gives it an Intelligence Index score of 15 and still ranks it first among open-weight models below 4B parameters. However, the two evaluations use different aggregation methods, prompts, and reasoning budgets, so their scores are not directly interchangeable. There is also no data yet from a tested API provider on throughput, time to first token, or cost.
There is also an important warning for deployment. Following the official `temperature=1.0, top_p=0.95` settings, one user tested the official Q8 and Q4 GGUF files on an RTX 3060 and reported that much of the output fell into loops when no repeat penalty was applied. Adding a penalty of approximately 1.15 produced a clear improvement. This is only a community result from a single machine and a single program, and the evaluation included repair passes, so it should not be treated as a universally optimal setting. Before adopting the model card defaults, engineering teams should test repetition rates, stopping conditions, tool-call success rates, and long-context memory usage with the actual quantization format they plan to deploy.