OpenBMB's MiniCPM5-1B Brings 128K-Context Reasoning to Your Smartphone
OpenBMB's MiniCPM5-1B packs a 128K context window, hybrid reasoning, and native tool-calling into a model that fits on a phone — and beats every competitor in its weight class.
Takeaways
- New SOTA at 1B scale: MiniCPM5-1B scores 42.57 avg across benchmarks, beating the next-best 1B competitor (LFM2.5-1.2B) at 35.61.
- Hybrid reasoning in one checkpoint: Toggle
enable_thinkingto switch between fast assistant and step-by-step reasoning mode without loading a different model. - 128K context on-device: Supports 131,072-token context and MCP-native tool calling — runs fully offline on consumer hardware including phones.
- RL + OPD training: Post-training with reinforcement learning and On-Policy Distillation raised scores by 16 points and cut overlong responses by 29 percentage points.
- Fully open stack: Weights, pre-training corpus (Ultra-FineWeb), and SFT data all released under Apache-2.0 — the recipe is reproducible, not just runnable.
- Available now: GGUF on Hugging Face (688 MB Q4_K_M), compatible with llama.cpp, Ollama, LM Studio, vLLM, and Apple Silicon MLX.
The conventional wisdom in AI has been that capable reasoning, long-context support, and agentic tool use belong to large models running in the cloud. MiniCPM5-1B from OpenBMB is a direct challenge to that assumption. At just over one billion parameters, it is the first release in the MiniCPM5 series , and it claims state-of-the-art performance within its size class while fitting inside a smartphone's memory budget.
One checkpoint, two modes
MiniCPM5-1B is a dense 1B Transformer built for on-device, local deployment, and resource-constrained scenarios. What makes it unusual is that it ships a single model file that can operate in two distinct modes. A built-in <think> chat template lets you switch via enable_thinking , the same checkpoint serves as both a fast assistant and a deliberate reasoner. Flip the flag and the model starts working through problems step-by-step before answering; flip it back and you get low-latency responses.
The model has 1,080,632,832 total parameters, 679,552,512 non-embedding parameters, 24 layers, grouped-query attention with 16 query heads and 2 key-value heads, and a context length of 131,072 tokens. That 128K context window , roughly 96,000 words , is remarkable for a model this size. MiniCPM5-1B supports the Model Context Protocol (MCP) and native tool calling, enabling local agent workflows on consumer devices without cloud connectivity.

Beating models twice its size
OpenBMB compares MiniCPM5-1B against strong open-source models in the same size class , LFM2.5-1.2B-Thinking, Qwen3-0.6B/think, and Qwen3.5-0.8B/think , and within this comparison set, MiniCPM5-1B reaches 1B-class open-source SOTA, with its advantage most visible in tool use, code generation, and difficult reasoning. OpenBMB reports an average score of 42.57 for MiniCPM5-1B Thinking across its selected benchmark set, compared with 26.77 for Qwen3-0.6B Thinking, 25.14 for Qwen3.5-0.8B Thinking, and 35.61 for LFM2.5-1.2B Thinking.