Zhidx reported on 8 September that Beijing edge-LLM unicorn ModelBest, together with the OpenBMB open-source community, has open-sourced MiniCPM5-2B, a next-generation edge language base model. The 2-billion-parameter model supports tool calling, deep search and code generation, and shipped day-zero adaptation to Intel, Rockchip and Arm platforms.
On the Artificial Analysis leaderboard MiniCPM5-2B scored 23 points to rank first among open models below 4 billion parameters, showing higher intelligence density than Qwen3.5 9B and the roughly six-times-larger Gemma 4 12B. On the Agentic Index it scored 20 against below 10 for peers its size.
Against a 1,000-point human baseline, MiniCPM5-2B reached 891 on real-task evaluation, far above Gemma 4 12B at 647, and across 34 benchmarks it averaged 53.9 to top its comparison group, beating models such as Qwen3.5-4B. ModelBest says the 2B scale now shows the first shape of general agent ability, and that edge-agent cost in office and life scenarios should keep falling.
Open recipe and training framework
Alongside weights, ModelBest and OpenBMB released the self-built reinforcement-learning framework Meshy and the reward-based strategy JustRL II, plus full training recipes, RL frameworks and datasets so developers can reproduce and verify. By August 2026 the MiniCPM series had passed 50 million global downloads.
Meshy drops the central controller and Ray dependency, organising training, inference and reward as peer services driven by data readiness through a TransferQueue, and switching between sync, async and full-async modes. JustRL II uses a three-stage pipeline to filter and calibrate data and adds a Critic for token-level credit assignment inside GRPO, lending the post-training a clear gain.
Day-zero chip adaptation
On Intel, the team optimised operators, graph fusion and memory scheduling on Core Ultra CPU, GPU and NPU with OpenVINO for local deployment. On Rockchip it runs on the RK3588 and RK1828 dual-chip platform through the RKNN3 toolchain with quantisation and operator optimisation. On Arm it adapted to second-generation Scalable Matrix Extension on Armv9 devices, where optimisation lifted prefill and decode performance to about 1.7 times and 1.2 times the baseline.
More images from the source report:
Editor’s note: This is an adapted translation of the original Zhidx report. It has been trimmed and restructured for readability for an international business audience.
Translated and adapted from Zhidx (https://www.zhidx.com/p/591922.html).