MiniCPM5-1B

MiniCPM5-1B is the opening model of the MiniCPM5 series from OpenBMB, the open-source arm of the Beijing company ModelBest. It is a dense transformer of 1,080,632,832 parameters — of which 679,552,512 sit outside the embedding tables — arranged in 24 layers with grouped-query attention, 16 query heads against 2 key/value heads. The published configuration gives a context window of 131,072 tokens and bfloat16 weights, and the architecture is the standard LlamaForCausalLM, so the model loads in ordinary tooling without custom code. The design target is stated plainly: local assistants, coding agents and tool-use workflows on hardware that cannot reach a data centre. The same checkpoint serves two behaviours through a built-in chat template — a direct answering mode and a deliberate one that emits an explicit reasoning block — selected by a single flag rather than by swapping models. Alongside the main release OpenBMB publishes the pre-training checkpoint, the supervised-tuning checkpoint, a GGUF build for llama.cpp, Ollama and LM Studio, and an MLX build quantised to 4 bits for Apple Silicon. Two things set this release apart from most small models. The first is the licence: Apache 2.0, with no revenue ceiling, no registration and no separate terms for commercial use — a contrast worth noting in a field where "open weights" increasingly comes with a threshold attached. The second is disclosure of the training material: the corpora used for pre-training and instruction tuning are released as public datasets rather than described in general terms, and the post-training recipe — supervised tuning, reinforcement learning, then on-policy distillation from domain-specific teacher models back into one released checkpoint — is documented step by step. The company's own evaluation compares the model against other releases in the same size class and claims the lead within that comparison set, with the largest gains in tool use, code generation and harder reasoning; those figures are the vendor's and have not been independently reproduced. What is measurable from outside is uptake: the repository records over 1.09 million downloads in the preceding thirty days and more than a thousand community endorsements.

▸Videos

No videos yet.

▸Reviews