#99th of 10 in Open models
minicpm-v
Open-weight multimodal models for image, video, document and speech understanding
Live demo ↗
(opens in a new tab)Documentation ↗
(opens in a new tab)Repository on GitHub ↗
(opens in a new tab)
- Stars
- 26.5k
- License
- Apache-2.0
- Last commit
- Sep 2026
- Last release
- May 2025
- Language
- Python
Missed the latest GitHub check; these facts may be older.
Overview
Repository for the MiniCPM-V 4.7 vision-language models (35B-A3B MoE and 1B) and MiniCPM-o 4.5, a 9B omnimodal model with full-duplex video, audio and speech output. It provides weights, inference and fine-tuning docs, and on-device deployment paths for iOS, Android and HarmonyOS. Code and weights are Apache-2.0.
Who it is for: Engineers running or fine-tuning small multimodal models locally
Strengths
- Two size classes: 1B and 35B-A3B MoE for vision, plus a 9B omnimodal model
- README reports 13+ tokens/s with 3 GB memory for 35B-A3B on iPhone 17 Pro Max
- Covers iOS, Android, HarmonyOS and Intel, AMD, Arm, Rockchip and Houmo chips
- Apache-2.0 for code and weights; fine-tuning via LLaMA-Factory and SWIFT, plus SRFT
Weaknesses
- No Docker or compose files; setup follows per-model docs
- Last tagged release is 2025-05-27, while model updates are announced in the README
- Benchmark claims are the authors' own, shown as images
- Hardware requirements for desktop or server use are not stated in the README
What it needs
- RAM ≥ 3 GB
- GPU optional
- Needs Hugging Face Transformers, llama.cpp-omni, Ollama
- Models: MiniCPM-V 4.7 35B-A3B, MiniCPM-V 4.7 1B, MiniCPM-o 4.5 9B, GGUF, BF16
Also in Open models
See all 10| Rank | Project | Score |
|---|---|---|
| 1 | llmfitRust tool that checks which LLMs fit your RAM, CPU and GPU | 83 out of 100 |
| 2 | layaSingle-pass multilingual classifier for typed choice, score and yes/no decisions | 76 out of 100 |
| 3 | timesfmPretrained time-series foundation model for zero-shot forecasting from Google Research | 70 out of 100 |
| #4 | unlimited-ocrOCR model that parses long documents and multi-page PDFs in one pass | 55 out of 100 |
| #5 | insightfaceFace detection, recognition and alignment toolbox with ArcFace and RetinaFace models | 52 out of 100 |