#1717th of 34 in Open models
minicpm
Small open-weight LLM family for on-device use, latest is MiniCPM5-2B
Live demo ↗
(opens in a new tab)Documentation ↗
(opens in a new tab)Website ↗
(opens in a new tab)Repository on GitHub ↗
(opens in a new tab)
- Stars
- 11.6k
- License
- Apache-2.0
- Last commit
- Sep 2026
- Last release
- May 2026
- Language
- Jupyter Notebook
8,671 on 31 Mar 202611,647 on 20 Oct 2026
From GitHub's daily star counts.
Checked on GitHub.
Overview
Weights and code for the MiniCPM series of compact dense LLMs. The current release, MiniCPM5-2B, is a 2.5B-parameter Llama-architecture model with a 131,072-token context, published in BF16, GGUF, MLX and GPTQ builds. The repo also links the training datasets (UltraData family), a DSpark speculative-decoding draft model, and deployment and fine-tuning agent skills.
Who it is for: Engineers running small LLMs locally or on edge devices
Strengths
- Standard LlamaForCausalLM architecture, so it loads in vLLM, SGLang, llama.cpp and Transformers
- Ships GGUF, MLX and GPTQ builds plus a DSpark draft model for speculative decoding
- Pretraining, code, SFT, agent and RL datasets are released alongside the model
- 131,072-token context in a 2B-class model; 1B and older sizes also available
Weaknesses
- Benchmark claims come from the authors' own comparison set; no independent results cited
- Repo is mostly docs and notebooks; weights are hosted on Hugging Face and ModelScope
- Needs separate inference software; no bundled server or chat UI
- Wiki is in Chinese only; RAM and VRAM requirements are not stated in the README
What it needs
- GPU optional
- Needs vLLM, SGLang, llama.cpp, Transformers
- Models: BF16, GGUF, MLX, GPTQ, MiniCPM5-2B
Also in Open models
See all 34| Rank | Project | Score |
|---|---|---|
| 1 | llmfitRust tool that checks which LLMs fit your RAM, CPU and GPU | 87 out of 100 |
| 2 | layaSingle-pass multilingual classifier for typed choice, score and yes/no decisions | 84 out of 100 |
| 3 | timesfmPretrained time-series foundation model for zero-shot forecasting from Google Research | 75 out of 100 |
| #4 | speechPyTorch toolkit for training and deploying ASR, TTS and speech LLM models | 70 out of 100 |
| #5 | yueLyrics-to-song model that writes an editable melody and chord score first | 65 out of 100 |