Member-only story
A 1B model walked into a ring with three giants. The 1B model did not leave.
MiniCPM5–1B is the first release in the new MiniCPM5 series from OpenBMB, the Tsinghua NLP lab. It is a 1,080,632,832-parameter dense Transformer that runs Apache-2.0, fits a 131K context window, and ships with native tool calling through SGLang. On the Artificial Analysis Intelligence Index, the reasoning variant scores 12, placing it #2 of 43 in the 1B-class open-weights tier. The non-reasoning variant is #1 of 29 in the non-reasoning 1B-class tier. The price per million tokens is $0.00 in, $0.00 out, because the weights are open.
This is the kind of release I have been waiting for as a builder, not as a benchmark-watcher. A 1B model that actually runs agentic tool use at a level I would route a real workflow through. A model that pulls the 1B cognitive-core idea (small, stripped-down, focused on reasoning + tool use + retrieval, the way Andrej Karpathy has been arguing for) into a real Apache-2.0 release with an active model card and an active training-data set.
Before I get into the numbers, the founder aside: I am Caspar Bannink. I write about AI model releases, agentic coding, and the kind of routing decision a builder has to make when a new model drops. I also build HomeScout, an AI…