OpenBMB's MiniCPM5-1B Beats Every Open 2B Model Running on Your Phone
OpenBMB's MiniCPM5-1B packs a 128K context window, hybrid reasoning, and native tool-calling into a 1B model that runs on-device — and tops the sub-2B open-weights leaderboard.
Takeaways
- New SOTA at 1B scale: MiniCPM5-1B scores 17.9 on the Artificial Analysis Intelligence Index, highest of any open-weights model under 1B parameters by 7.4 points.
- Beats larger models: Outperforms Alibaba's Qwen3.5-2B (16.3) at less than half the parameter count, with an average benchmark score of 42.57 vs. 35.61 for the next-best sub-2B model.
- Novel training recipe: Post-training combines 400B tokens of SFT, two-stage RL with specialized teachers, and On-Policy Distillation (OPD) — yielding a 16-point benchmark gain and 29pp reduction in runaway-length responses.
- Hybrid reasoning, single checkpoint: Toggle between fast assistant mode and deliberate chain-of-thought reasoning via
enable_thinkingflag, with native tool-calling and MCP support. - Runs anywhere: Available as BF16, GGUF (Ollama/llama.cpp/LM Studio), and MLX/4-bit for Apple Silicon; supports 9+ AI chip vendors via FlagOS.
- Fully open: Weights, training corpus (Ultra-FineWeb, UltraData-Math, UltraData-SFT), and deployment cookbooks all released under Apache 2.0.
MiniCPM5-1B, the first release in OpenBMB's fifth-generation MiniCPM series, is a 1.08-billion-parameter language model designed to run entirely on-device. OpenBMB is a China-based lab jointly founded in 2022 by Tsinghua University's NLP Lab and ModelBest Inc. The headline claim is hard to ignore: MiniCPM5-1B scores 17.9 on the Artificial Analysis Intelligence Index, the highest of any open-weights model at 1B parameters or below by 7.4 points. That puts it ahead of models with twice the parameter count.
A 1B model that beats 2B competitors
MiniCPM5-1B reaches an average score of 42.57 across reasoning, knowledge, code, instruction-following, math, logic, and agentic benchmarks, above the highest average score of 35.61 among strong open-source models in the same size class. The comparison set includes Alibaba's Qwen3-0.6B, Qwen3.5-0.8B, and Liquid AI's LFM2.5-1.2B-Thinking. It surpasses its predecessor MiniCPM-V 4.6 1.3B (12.7) by 5.3 points at ~23% fewer parameters, and beats Qwen3.5 2B (Reasoning, 16.3) by 1.6 points at less than half the parameter count.

One standout from independent testing: MiniCPM5-1B scores -1 on AA-Omniscience, the highest in its size class, earned by abstaining rather than hallucinating. Sub-2B peers typically attempt a large proportion of questions and hallucinate at high rates; MiniCPM5-1B declines the majority of questions, an honest posture that AA-Omniscience credits positively. Peers in the same size class score between -70 and -89 on the same metric.
Token efficiency is also notable. MiniCPM5-1B is more token-efficient than the larger reasoning peers it surpasses. It used 12.6M output tokens to run the Intelligence Index, ~31x fewer than Qwen3.5 2B (Reasoning, 389M) and ~8x fewer than Qwen3.5 2B (Non-reasoning, 100M).
Where it shines , and where it doesn't
It is designed for local assistants, coding agents, tool-use workflows, and reasoning scenarios where a compact model is preferred. Its strongest domains are agentic tool use, code generation, and competition-level math. MiniCPM5-1B can run local agents on phones and shows strengths in agentic tool use and code generation, but falters on logic traps. On-device agentic workflows at the 1B-parameter scale are now feasible, but reliability limits in complex reasoning mean practitioners should treat outputs as opportunistic rather than authoritative.