MiniCPM5-2B

MiniCPM5-2B is a small dense model that punches well above its size. It averages 53.9 across 34 benchmarks, ahead of the 4 billion parameter Qwen3.5-4B, and leads independent rankings of open-weight models under 4B. Its strongest results are practical ones: tool calling and code rather than trivia.

It carries a native context of 131,072 tokens, a hybrid thinking mode for harder questions, and an Apache 2.0 licence that allows commercial use. It runs in llama.cpp, Ollama, LM Studio, MLX, vLLM and SGLang, which makes it one of the easier capable models to try on hardware you already own.