OpenBMB Drops MiniCPM5-2B Open Weights Under Apache 2.0 for On-Device Agents

OpenBMB published MiniCPM5-2B open weights on Hugging Face without a launch event: the openbmb/MiniCPM5-2B repository appeared September 6, 2026, and OpenBMB’s changelog marks the release September 7 under Apache 2.0. OrcaRouter’s September unpack notes ungated BF16 weights plus GGUF, MLX, GPTQ, and DSpark companions, with UltraData training corpora released alongside the model. Artificial Analysis lists the model as released September 7, 2026, with Apache 2.0 open weights and a 131k-token context window.

MiniCPM5-2B is the second MiniCPM5 series model after MiniCPM5-1B. The Hugging Face model card describes a dense LlamaForCausalLM with 2,516,756,480 parameters (1,981,982,720 non-embedding), 42 layers, grouped-query attention with 16 query heads and 2 KV heads, and max_position_embeddings set to 131,072. OrcaRouter flags that some World Artificial Intelligence Conference materials had pitched a 512K context story; as shipped, the checkpoint documents 131,072, and Artificial Analysis likewise reports a 131k context window.

Plain Llama stack, agent-oriented post-training

Because the architecture is standard Llama, the card says engines need no custom kernels or model-code fork. Documented quickstarts cover vLLM 0.21 or newer, SGLang 0.5.16 or newer, and Transformers 5.6 or newer. For tool calling, OpenBMB recommends SGLang with its built-in minicpm5 parser, which converts the model’s XML-style tool calls into OpenAI-compatible tool_calls. A companion DSpark draft model is documented for speculative decoding in SGLang without changing target outputs.

Post-training follows SFT, specialized RL teachers, and On-Policy Distillation (OPD). On the model card’s listed benchmarks, OpenBMB credits RL + OPD with average gains of 10.96 points on reasoning and general capabilities and 6.96 points on agentic capabilities—figures the lab presents as internal deltas. Agent-oriented training data is released as UltraData-SFT-Agent-2609 and UltraData-RL-2609 within the open UltraData family.

Benchmarks: separate internal rows from AA composites

The model card’s comparison-set average of 53.9 (versus Qwen3.5-4B at 51.1 and other 2B–4B open models) is an OpenBMB-normalized score; many headline rows are “reproduced internally,” while dagger-marked scores come from Artificial Analysis. OrcaRouter reports that Artificial Analysis Intelligence Index v4.2 placed MiniCPM5-2B at 15, the top open-weights result under 4B in that write-up. The live Artificial Analysis model page currently shows Intelligence Index 14 under v4.3—a newer methodology that must not be conflated with the v4.2 figure.

Primary sources for this brief are the Hugging Face openbmb/MiniCPM5-2B model card, OrcaRouter’s MiniCPM5-2B open-weights unpack, and Artificial Analysis’s MiniCPM5-2B page.