開放模型／端側推論
MiniCPM5-2B Releases Full Training Stages and On-Device Formats, Distilling 16 RL Experts into a Single Model
OpenBMB has released MiniCPM5-2B, its training data, and multi-stage weights spanning pretraining through RL, along with GGUF, MLX, GPTQ, and speculative decoding variants. The model retains the standard Llama architecture and a 128K context window, but the SGLang parser required for tool calling may still need to be installed from the main branch.
OpenBMB released the MiniCPM5-2B series on September 13. The “2B” in the name refers to approximately 1.982 billion non-embedding parameters; the full model has 2,516,756,480 parameters and 42 layers. It uses grouped-query attention with 16 query heads and two KV heads, and has a native context length of 131,072 tokens. It retains the standard `LlamaForCausalLM` architecture, so Transformers, vLLM, and SGLang do not need to load custom model code.
What stands out in this release is not just the final checkpoint. The team also provides Base, Midtrain, SFT, and RL+OPD versions, as well as BF16, GGUF, MLX 4-bit, GPTQ 4-bit, LiteRT, and images for multiple chip platforms, allowing researchers to trace which capabilities emerge at each training stage. OPD compresses 16 RL-trained expert models back into a single student model, five of which target agentic tasks. During distillation, it uses the reverse KL divergence between the teachers’ and student’s full vocabulary distributions as a per-token advantage signal, eliminating the need to create a separate distillation dataset. According to the project team, RL+OPD delivers an average gain of 10.96 points in reasoning and general capabilities, and 6.96 points in agentic capabilities.
The model card reports an overall average score of 53.9, above the highest-scoring 4B-class model in the same table, at 51.1. It scores 69.1 on LiveCodeBench v6 and 86.5 on AIME 2025. However, these are evaluations conducted by the development team, and the overall average combines tasks and metrics with different scales. They therefore cannot replace testing of on-device latency, memory usage, and real-world tool-call success rates.
There is also a deployment compatibility pitfall: the model emits XML-style tool calls. Although the documentation says SGLang includes a `minicpm5` parser, the bundled deployment skill notes that the pip release may not yet contain it, requiring installation from SGLang’s main branch. Engineering teams should pin the runtime commit and separately validate memory usage at 128K context length and the real-world acceptance rate of DSpark speculative decoding.