MiniCPM5-2B: the MiniCPM5 recipe at 2.5B, with every local format shipped on day one

OpenBMB put MiniCPM5-2B on Hugging Face on September 6, 2026 under Apache 2.0: 2,516,756,480 parameters (1.98B without embeddings), a standard `LlamaForCausalLM`, 131,072-token context, English and Chinese. By September 15 it had 206,774 downloads, and the official GGUF repo another 108,471 — the second most-downloaded new model in our window after DeepSeek-V4.1-Flash. Our September 7 sweep did not catch it; it was one day old and it had been released beside much louder news.

Verdict: OpenBMB scales its 1B recipe to 2.5B: Apache 2.0, 128K context, official GGUF and MLX, and 200K+ downloads in nine days — a small model we missed on September 7

The take

What it is, from the card. The second model in the MiniCPM5 series, scaling up the same training recipe as MiniCPM5-1B (May 22): SFT, then RL, then on-policy distillation. 42 layers, grouped-query attention with 16 query heads and 2 KV heads. OpenBMB released the whole ladder of checkpoints — Base, Midtrain, SFT-only and final — plus official GGUF (llama.cpp / Ollama / LM Studio), MLX 4-bit, GPTQ, a LiteRT build, and a small DSpark draft model for faster decoding. It also published the training datasets behind it under its UltraData family. That is more openness than most models ten times its size.

OpenBMB's claims, labelled as such: "2B-class open-source SOTA" within its own comparison set, with an average of 53.9 that it says also beats every larger model in that set (the highest is 51.1), and particular strength in code, maths, long context and tool use. These are vendor numbers on a comparison set OpenBMB chose. We have no independent measurement yet.

Where it fits. At about 1.5–2 GB at Q4 it runs on anything, including CPU-only boxes and phones. That is the same band as our low-tier picks, and those picks stay: Qwen 3.5 4B, Phi-4 Mini and Gemma 3 4B are proven, and a nine-day-old vendor benchmark does not displace them under our own rule. It is the natural candidate to watch for the smallest tier if independent numbers agree with the card. Two practical notes from the card: there is no Ollama library tag, so pull the official GGUF from Hugging Face; and in llama.cpp the default `min_p=0.05` can cause repetitive output, so OpenBMB recommends changing it.

Our call: no new model page. `/models/minicpm5-1b/` now covers the 2B sibling, and no planner pick changes.

Where this fits

Models: MiniCPM5-1B (Apache 2.0, OPD-trained) · Qwen 3.5 4B · Phi-4 Mini

Hardware: Mac Mini M4 16 GB · NVIDIA RTX 3060 12 GB

Sources

Next step

Try this in the planner→