BAAI's MiniCPM5-2B tops open-source rankings, brings generalist Agent skills on-device at 2B

机器之心 · wechat · 2026-09-08

Modelbest (OpenBMB) open-sourced MiniCPM5-2B, a 2B-parameter on-device model that natively supports tool calling, deep search, and code generation — an early blueprint for generalist agents running locally. Unusually, the release also includes training recipes, the Meshy RL framework, the JustRLII algorithm, and several datasets.

Benchmarks

Ranked #1 open-source model on Artificial Analysis Intelligence Index (23), ahead of larger Gemma 4 12B and Qwen3.5 9B
#1 on agent-workflow evals (20 vs. Granite 4.2 8B's 9); 21K output tokens per task on average
53.9 average across code/math/tool-use/search-agent benchmarks

Open tech and data

Meshy drops RL training's central controller and Ray dependency, flexibly switching sync/async/fully-async modes
JustRLII adds a Critic to GRPO for token-level credit assignment, plus a three-stage data pipeline, fixing the score-regression problem of long-CoT RL on small models
UltraX refines corpora with a 0.6B model predicting structured edit ops (keep/delete/replace/insert); a 1B model trained on 16B refined tokens beats 20B raw tokens, topping HF Datasets Trending
Open datasets: UltraData-Code (L0–L3 tiers, 400B/150B tokens), 500K agent SFT samples, 80K+ RL samples

Deployment: supports vLLM, llama.cpp, Ollama and the in-house CPU framework Arclight. Hands-on tests (meeting-notes QA, local résumé extraction, webpage generation) work, targeting privacy-sensitive on-device agent scenarios. Cloud-edge collaboration is framed as the emerging default.

Original post →

BAAI's MiniCPM5-2B tops open-source rankings, brings generalist Agent skills on-device at 2B

More from coding & agent