MiniCPM5-1B tops sub-2B open models with 17.9 Artificial Analysis score
OpenBMB released full open-source weights, training data, and deployment code for MiniCPM5-1B, a 1-billion-parameter model that outscores Qwen3.5-2B despite half the parameter count.
OpenBMB released MiniCPM5-1B on May 25 with full open-source weights, training data, and deployment code. The 1-billion-parameter model scored 17.9 points on Artificial Analysis, claiming first place among all open models under 2 billion parameters and beating Qwen3.5-2B's 16.3 points with half the parameter budget.
The efficiency gain is substantial in a category where parameter count typically drives performance. MiniCPM5-1B outperforms Qwen3.5-0.8B and LFM2.5-1.2B-Thinking across knowledge retrieval, mathematical reasoning, code generation, and tool-use tasks. An INT4-quantized version compresses to 0.5 GB, small enough to run on mobile phones, in-browser via WebAssembly, and on edge devices without server round-trips.
OpenBMB trained MiniCPM5-1B using ForgeTrain, a pretraining framework the team describes as the first production-grade LLM training system written entirely by AI without human programmers. The framework reportedly runs 10 percent faster than NVIDIA Megatron on equivalent hardware, though the announcement does not specify cluster size, GPU type, or whether the comparison used identical hyperparameters and data pipelines. The full-precision weights and training data are available on ModelScope, along with deployment scripts for common runtimes. If ForgeTrain's speed claim holds across different configurations and scales to larger models, it could reshape how teams approach infrastructure—but the next milestone will be whether the framework itself becomes open-source and whether efficiency gains translate beyond the 1B parameter range.
