MiniCPM5-2B is OpenBMB's on-device, 2.5-billion-parameter open-source language model, built for developers who need real agent and coding ability on phones, laptops, and edge devices. Released September 7-8, 2026, it is the second drop in the MiniCPM5 family after the 1B version and ships under an Apache-2.0 license.
Core Features
- A 131,072-token context window (128K) that fits long documents, full codebases, and multi-turn conversations.
- Native tool calling, deep search, code generation, and multi-step reasoning despite the small size.
- Standard LlamaForCausalLM layout with 42 layers and grouped-query attention, so vLLM, SGLang, llama.cpp, Ollama, and MLX load it without custom kernels.
- INT4 quantization shrinks the model to about 1.2GB while keeping 95% of full-precision quality.
- An Agentic Index score of 20, leading its size class on autonomous task execution.
Use Cases
- On-device assistants that must call tools and hold long context without a data-center GPU.
- Edge and robotics applications with tight memory and power budgets.
- Local coding agents and document-processing pipelines that keep data private.
- Researchers fine-tuning a transparent, fully open training stack.
Pricing
MiniCPM5-2B is free and open-weight under Apache-2.0, so the effective license cost is $0. OpenBMB published not just the weights but the full training recipe, datasets, and RL infrastructure, so there is no license fee or usage cap. The MiniCPM family has passed 50 million downloads, and deployment costs only the compute you run locally or in a lightweight cloud instance.
Our Take
Best for builders who want strong agent behavior on consumer hardware without cloud bills. The trade-off is weaker knowledge-heavy scores than 4B peers, so treat it as a compact coding and tool-use companion rather than a general knowledge engine. See related AI Models and Engines.
