MiniCPM5-2B

Second model in the MiniCPM5 series, released September 7, 2026, scaling the MiniCPM5-1B recipe to a dense 2.52B-parameter (1.98B non-embedding) standard LlamaForCausalLM: 42 layers, hidden size 2,048, 16 query and 2 KV heads, a 130K vocabulary, and a native 131,072-token context. Hybrid thinking mode (on or off per request), Apache 2.0, and no custom kernels, so vLLM, SGLang, and llama.cpp load it directly; OpenBMB ships Base, Midtrain, SFT, GGUF, MLX, and GPTQ checkpoints plus a DSpark draft model for speculative decoding, and FlagOS ports to nine chip families. The pretraining corpus (Ultra-FineWeb, Ultra-FineWeb-L3, UltraX, UltraData-Code, UltraData-Math) is released; the pretraining token count is not stated.

Training follows OpenBMB's UltraData tiered-data pipeline: base training (stable then decay), mid-training, then a three-step post-training of 400B tokens of deep-thinking SFT, critic-based RL per the JustRL II recipe that scales small models to 128K-token reasoning, and on-policy distillation that folds 16 RL expert models (5 of them agentic) into one release model using full-vocabulary reverse KL as the advantage. OpenBMB credits RL plus OPD with +10.96 points on reasoning and general benchmarks and +6.96 on agentic ones. Against its own 2B and 4B comparison set it averages 53.9 versus 33.2 for LFM2.5-2.6B, 28.0 for Qwen3.5-2B, and 51.1 for the best larger model (Qwen3.5-4B): LiveCodeBench v6 69.1, AIME 2025 86.5, MATH-500 94.6, MMLU-Pro 70.8, GPQA Diamond 70.2, IFBench 66.3, BFCL v4 66.6, SWE-bench Verified 46.4, AA-LCR 59.0, NoLiMa 68.1, and BrowseComp-ZH 43.5; it trails Qwen3.5-4B on SWE-bench Pro (14.4 vs 28.2) and Terminal-Bench 2.1 (8.6 vs 25.8). AA Intelligence Index v4.3: 12, which AA notes is well above the median of 6 for open-weight models of similar size.