OpenBMB's MiniCPM5-1B Beats Every Open 1B Model by 7 Points
OpenBMB's MiniCPM5-1B packs a 128K context window, hybrid reasoning, and native tool-calling into a 1B model that runs entirely on-device
Takeaways
- New SOTA at 1B scale: MiniCPM5-1B scores 42.57 avg across 7 benchmark categories, beating all open-source sub-2B models.
- Runs fully on-device: Fits in ~0.5GB at 4-bit; supports phones, Apple Silicon (MLX), CPU, and 9+ chip families via FlagOS.
- 128K context + tool calling: Native MCP and XML-style function calling, with SGLang converting to OpenAI-compatible format.
- Hybrid reasoning in one checkpoint: Toggle between fast assistant mode and step-by-step thinking mode via
enable_thinking. - RL + On-Policy Distillation: Post-training boosted scores by +16 points and cut runaway-length responses by 29 percentage points.
- Fully open stack: Weights, pre-training corpus (Ultra-FineWeb), and SFT data all released under Apache 2.0.
MiniCPM5-1B, the first release in OpenBMB's MiniCPM5 series, is a 1.08-billion-parameter language model built from the ground up for on-device deployment. It runs on phones, laptops, and Apple Silicon, fits in roughly 0.5GB at 4-bit quantization, and requires no cloud API. For a model this small, the benchmark numbers are genuinely surprising.
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. Its nearest competitors are Alibaba's Qwen3-0.6B and Qwen3.5-0.8B, and Liquid AI's LFM2.5-1.2B-Thinking. Independent analysis from Artificial Analysis confirms the lead: 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. 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 checkpoint, two modes
MiniCPM5-1B has a built-in <think> chat template, switchable via enable_thinking. The same checkpoint serves as both a fast assistant and a deliberate reasoner. This is more than a cosmetic toggle. In thinking mode, the model works through problems step by step before answering; in no-think mode, it responds immediately. The recommended sampling parameters differ between the two modes:
- Think mode:
temperature=0.9, top_p=0.95withenable_thinking=True - No Think mode:
temperature=0.7, top_p=0.95withenable_thinking=False
The model has 1,080,632,832 total parameters, 679,552,512 non-embedding parameters, 24 layers, grouped-query attention (GQA) with 16 query heads and 2 key-value heads, and a context length of 131,072 tokens. That 128K context window is a meaningful edge for an on-device model. The 128K-token context equals roughly 96,000 words of continuous text, which OpenBMB says enables long roleplay sessions, full-document digests, and agent contexts that persist across many exchanges.