MiniCPM

weights

open(Apache-2.0, ungated on HF across generations 3-5)

data

closed(training-data research published (Ultra-FineWeb, UltraClean) but the corpus is not released end to end)

code

partial(OpenBMB/MiniCPM is inference/deployment plus papers

license

Apache-2.0(OSI)

Apache-2.0 weights, ungated in every SKU read across generations 3 through 5, with training-data research published as papers and partial dataset artifacts but no end-to-end corpus release, and a repository that carries inference and deployment rather than the pretraining pipeline. Open weights, closed recipe: 3 by the fallthrough, governed by the current MiniCPM5 generation.

• https://api.github.com/repos/OpenBMB/MiniCPM recorded 2026-09-01

  `license: Apache-2.0`, 10,288 stars, description "MiniCPM5-1B: A SOTA 1B on-device LLM, small yet powerful." — confirms MiniCPM5 is the current lead release; the repo is model usage/inference, not a training pipeline.

• https://huggingface.co/api/models/openbmb/MiniCPM5-1B recorded 2026-10-07

  Model API for openbmb/MiniCPM5-1B: `gated: false`, `private: false`, cardData.license `apache-2.0`, safetensors weights listed; the card names Ultra-FineWeb, Ultra-FineWeb-L3, UltraData-Math and UltraData-SFT-2605 as its datasets.

• https://huggingface.co/openbmb/MiniCPM4.1-8B/raw/main/README.md recorded 2026-10-07

  Raw card: `license: apache-2.0`; links the GitHub repo and technical report (arXiv 2506.07900) and the InfLLM-V2 paper; a hybrid reasoning model card with no released training corpus and no pretraining pipeline link.

Adoption is measured as Hugging Face downloads summed across the four declared checkpoints spanning MiniCPM3 through the current MiniCPM5 generation. MiniCPM5-1B, the newest release, accounts for most of that total.

• https://huggingface.co/api/models/openbmb/MiniCPM3-4B recorded 2026-09-01

  26,614 downloads in the trailing 30 days for openbmb/MiniCPM3-4B

• https://huggingface.co/api/models/openbmb/MiniCPM4-8B recorded 2026-09-01

  37,939 downloads in the trailing 30 days for openbmb/MiniCPM4-8B

• https://huggingface.co/api/models/openbmb/MiniCPM4.1-8B recorded 2026-09-01

  47,819 downloads in the trailing 30 days for openbmb/MiniCPM4.1-8B

• https://huggingface.co/api/models/openbmb/MiniCPM5-1B recorded 2026-09-01

  810,294 downloads in the trailing 30 days for openbmb/MiniCPM5-1B

MiniCPM is competitive with same-size Qwen3 and Gemma checkpoints and is unusually efficient at long context via sparse attention, but the ceiling is the size class itself, well below the category's open-weight frontier. It is level with Smollm, the category's other small-model line, which makes the same claim: outperforming same-size peers while competitive one size up.

• https://huggingface.co/openbmb/MiniCPM4.1-8B/raw/main/README.md recorded 2026-10-07

  The 4.1-8B card presents the line as "a hybrid reasoning model with trainable sparse attention," usable in deep-reasoning and non-reasoning modes, with benchmark comparisons against similar-size open models.