MiniCPM

Small enough for phones and laptops, the MiniCPM models from OpenBMB aim for top quality at roughly 1 to 8 billion parameters. The newest MiniCPM5 checkpoints, a 1B released in May 2026 and a 2B in September 2026, use the standard Llama architecture with a 131K-token context window, so llama.cpp, Ollama, vLLM, SGLang, MLX, LM Studio and Hugging Face Transformers load them without custom code, with official GGUF and MLX builds. In the team's own tests the 2B beats similar-size open models on average, notably in code, math, long context, tool use and agents. Earlier branches remain available, including MiniCPM4 and 4.1 with trainable sparse attention, MiniCPM-SALA for million-token context, and ternary BitCPM variants; vision models live in the separate MiniCPM-V repository. The repo adds step-by-step guides for serving and for tuning with TRL, LLaMA-Factory, ms-swift and Unsloth, and OpenBMB has published much of the training data. Code and weights are Apache 2.0, which permits commercial use. A 2B model runs on CPU through llama.cpp or on a modest GPU, a fit for offline assistants, embedded agents and cheap fine-tuning tests.