OpenBMB Open-Sources MiniCPM5-2B, Topping Sub-4B Open Models

On September 7, OpenBMB and ModelBest released and open-sourced MiniCPM5-2B, a compact on-device model with roughly 2.6B parameters. Per Artificial Analysis's official evaluation, it scores 15 on Intelligence Index v4.2 — the highest among open-weight models of 4B parameters or fewer — and 20 on the Agentic Index. Weights are available on Hugging Face under Apache 2.0, alongside a GPTQ quantized build, Day-0 support in vLLM v0.21.0, and open training recipes, the Meshy RL framework, the JustRLII algorithm and datasets.

Confirmed

• MiniCPM5-2B is a 2B dense model built on the standard LlamaForCausalLM architecture with native 128K context (confirmed by @vllmproject)
• Scores 15 on Artificial Analysis Intelligence Index v4.2, ranking first among sub-4B open-weight models (confirmed by @ArtificialAnlys)
• Agentic Index of 20, with tool calling, deep search and code generation; ModelBest claims 10x agent capability over same-class peers
• Apache 2.0 weights on Hugging Face, plus a GPTQ quantized version in safetensors for on-device/edge deployment
• Day-0 support in vLLM v0.21.0 (officially confirmed)
• Open-sourced training recipes, Meshy RL framework, JustRLII algorithm and datasets (reported by @机器之心, @AGI Hunt)

Unconfirmed

• Multiple accounts, including ModelBest itself, cite an intelligence score of 23, conflicting with Artificial Analysis's official 15; the official figure of 15 takes precedence
• ModelBest's claim of surpassing Gemma 4 12B and Qwen3.5-9B is a vendor assertion without third-party verification

Why it matters

• It pushes intelligence density leadership into sub-4B on-device models while bringing agentic capabilities to a 2B model, lowering the barrier to local deployment
• Full-stack openness — Apache 2.0, GPTQ, vLLM Day-0, recipes, RL framework and datasets — forms a ready-to-use and reproducible inference and research ecosystem

2026-09-07 ~ 2026-09-08 · 14 related posts

• [source] OpenBMB ships MiniCPM5-2B GPTQ for on-device long-context inference — openbmb · 2026-09-07
• [source] OpenBMB's MiniCPM5-2B tops sub-4B open models with Intelligence Index score of 15 — ArtificialAnlys · 2026-09-07
• OpenBMB releases Apache 2.0 MiniCPM5-2B, topping its class on AA Intelligence Index — ArtificialAnlys · 2026-09-07
• [source] MiniCPM5-2B open-sourced with Day-0 vLLM support, tops sub-4B open models — vllm_project · 2026-09-07
• OpenBMB ships MiniCPM5-2B, top open-weights score under 4B params — Equivalent-Grass-527 · 2026-09-07
• OpenBMB open-sources MiniCPM5-2B, topping the sub-4B open-model intelligence index — LearnWithBishal · 2026-09-08
• MiniCPM5-2B open-sourced: top sub-4B model with 10x agentic lead, full data and RL recipe released — 面壁智能 · 2026-09-08
• BAAI's MiniCPM5-2B tops open-source rankings, brings generalist Agent skills on-device at 2B — 机器之心 · 2026-09-08
• ModelScale open-sources MiniCPM5-2B with weights, data, RL framework — 2B beats 4B rivals — AGI Hunt · 2026-09-08
• MiniCPM5-2B: a 2B open model tops sub-4B leaderboard with 10x agent score, recipe fully open-sourced — 新智元 · 2026-09-08
• OpenBMB open-sources MiniCPM5-2B, tops intelligence index for sub-4B models — Aiden_Tech_Ai · 2026-09-08
• ModelBest open-sources MiniCPM5-2B, topping sub-4B open models with full training recipe — 智东西 · 2026-09-08

2 near-duplicate retellings: hey_abusiddik · hey_abusiddik