MiniCPM5-2B: a 2B open model tops sub-4B leaderboard with 10x agent score, recipe fully open-sourced
新智元 · wechat · 2026-09-08
ModelBest and OpenBMB open-sourced MiniCPM5-2B, which tops the Artificial Analysis index for open base models under 4B params (23 points), beating Gemma4 12B and Qwen3.5 9B. Its AA Agentic Index score of 20 is roughly 10x that of typical same-size models (2 points), and it scored 891 on GDPval-AAv2, closest to the human baseline.
Technical details
- Only 21k output tokens per task, the lowest among peers — faster and cheaper on-device inference.
- Fully open: training recipe, the Meshy RL framework (no central controller, sync/async/full-async modes), and four datasets — UltraX row-level web data refinement (20% training savings), UltraData-Code, 500k+ agent SFT samples, and a three-stage-cleaned RL set.
- JustRL-II adds a Critic to GRPO for token-level credit assignment.
- Compatible with LLaMA-Factory, vLLM, llama.cpp, Ollama, and more.
ModelBest has now open-sourced 10+ training datasets over three years (2.66M+ downloads), arguing that open data pipelines matter more to the field than open weights as compute gains plateau.
Related event: OpenBMB Open-Sources MiniCPM5-2B, Topping Sub-4B Open Models(14 posts)→
More from Models
- Open Models Match GPT-6 on MedMCQA but Trail 17 Points on Hard Clinical Cases — keunwoochoi · 2026-10-27
- Health-Focused Benchmark Reranks LLMs: Muse Spark Open Jumps from #9 to #1 — keunwoochoi · 2026-10-27
- DeepSeek cut KV cache per token from 389KB to 890B in two years as open-closed gap debate rages — teortaxesTex · 2026-10-27
- GPT-6 Astra controls dexterous hands zero-shot at just 5.2% success in 500 trials — stingning · 2026-10-27
- Unverified: fable 5.5 speculated at 4-5T parameters to outscale 1-3T Opus — haider1 · 2026-10-27
- Kimi K3 Is 'Open' Until You Succeed: License Caps Free Use at $20M Revenue — alex_verem · 2026-10-27