OpenBMB Releases MiniCPM5-2B Under Apache 2.0 With a 131K Context and the Top Intelligence Index Score for Any Open Model Under 4B

OpenBMB / Hugging Face·medium signalXBlueskyLinkedInCopy link

OpenBMB published MiniCPM5-2B on September 7, a 2B-parameter dense reasoning model with 42 layers, grouped-query attention, BF16 weights, a 131,072-token context and a LlamaForCausalLM architecture, licensed Apache 2.0 on Hugging Face. It scores 15 on Artificial Analysis Intelligence Index v4.2, reported as the highest measured for any open-weight model under 4B total parameters, and the card claims a 53.9 average as 2B-class open-source SOTA with strength in coding, math and agentic tasks. Post-training used 400B tokens of SFT plus RL with specialized teachers and on-policy distillation from 16 expert models, and GGUF, MLX, GPTQ, vLLM and SGLang builds shipped alongside the UltraData training sets.