MiniCPM5-2B open-sourced: top sub-4B model with 10x agentic lead, full data and RL recipe released
面壁智能 · wechat · 2026-09-08
ModelBest and OpenBMB open-sourced MiniCPM5-2B, a 2B on-device model that tops the AA Index among sub-4B open base models (23 points, beating Gemma 4 12B and Qwen3.5-9B) and scores 20 on the Agentic Index — 10x its 2-3B peers — supporting tool calling, deep search and code generation for on-device agents.
The release bundles four datasets (UltraData-Code 550B tokens, 500k SFT-Agent samples, 80k RL samples, and the 100B-token UltraX refinement framework), the Ray-free RL training framework Meshy, and JustRL II, a GRPO-with-Critic algorithm for token-level credit assignment. Day0 ports cover Intel Core Ultra, Rockchip RK3588+RK1828, and Arm SME2 devices (prefill 1.7x, decode 1.2x faster). MiniCPM family downloads have surpassed 50M.
More from Models
- Repeating instructions in prompts still makes models adhere better, like it's 2023 — tokenbender · 2026-09-08
- Model slips on a science knowledge test, author calls it acceptable — felpix_ · 2026-09-08
- Native video inference beats sampled frames, but OpenRouter providers don't support it yet — spillai · 2026-09-08
- Users game Tibo's token reset: spend 80% fast, stagger reset cycles to dodge surprise resets — tinyfool · 2026-09-08
- Economists in the top 10% of AI use see no step change from GPT 5.6 to 6 — aniketapanjwani · 2026-09-08
- Mathematician: Astra solved two of my unpublished theorems in ~60 hours each, $20k in API — basedjensen · 2026-09-08