@OpenBMB dropped MiniCPM5-2B three days ago and it is the best open model under 4B right now. 2.5B params, Apache 2.0, 128K context, and it beats Qwen3.5 4B with 44 percent fewer parameters. Number 1 open weights model under 4B on the Artificial Analysis Intelligence Index, level with Qwen3.5 9B at roughly a quarter of the size. Where it really stands out is code and agents. On OpenBMB's own table it averages 53.9 across 34 benchmarks against 51.1 for the best 4B model. LiveCodeBench 69.1 vs 56.4 for Qwen3.5 4B. SWE-bench Verified 46.4 vs 33.6. Tool use at 97 on tau2 Telecom. And it is one checkpoint that can think or answer fast, you pick per request. The trick is how they built it. They RL trained 16 specialist models, 5 of them agentic, then distilled all of them back into this one 2B model. Full training data is public too. Just ran it on the @RunAnywhereAI Android app btw, llama.cpp, GGUF pulled straight from Hugging Face inside the app, nothing leaves the phone.