MiniCPM-V
OpenBMBMiniCPM-V is OpenBMB's line of small vision-language models built to run on phones and other edge devices. The models read single images, several images at once and video, and the current MiniCPM-V 4.6 pairs a SigLIP2-400M encoder with a Qwen3.5-0.8B language model and supports tool calling. The line is separate from OpenBMB's text MiniCPM models and shares its repository with the omni-modal MiniCPM-o.
Openness
3 high confidence- weights
- open(safetensors on the Hub, ungated)
- data
- partial(the RLAIF-V preference set is released under CC-BY-NC-4.0
- code
- partial(official fine-tuning scripts for users (finetune/)
- license
- Apache-2.0(OSI
MiniCPM-V 4.6 and its code are Apache 2.0, and OpenBMB ships fine-tuning scripts. It releases its RLAIF-V preference data, under a non-commercial license, but not the rest of the training mixture or its own training recipe.
- https://cdn.jsdelivr.net/gh/OpenBMB/MiniCPM-V@main/finetune/readme.md recorded 2026-09-27
"We offer the official scripts for easy finetuning of the pretrained MiniCPM-V 4.0, MiniCPM-o 2.6, MiniCPM-V 2.6 ... on downstream tasks"
- https://cdn.jsdelivr.net/gh/OpenBMB/MiniCPM-V@main/LICENSE recorded 2026-09-27
The repository LICENSE is the Apache License, Version 2.0
- https://huggingface.co/api/datasets/openbmb/RLAIF-V-Dataset recorded 2026-09-27
RLAIF-V-Dataset: public, ungated, cc-by-nc-4.0; "also used in MiniCPM-V 4.5"
- https://huggingface.co/api/models/openbmb/MiniCPM-V-4_6?expand[]=downloads&expand[]=cardData&expand[]=gated&expand[]=createdAt&expand[]=lastModified&expand[]=siblings recorded 2026-09-27
Hub metadata for MiniCPM-V-4.6: gated false, model.safetensors, card license apache-2.0
- https://huggingface.co/openbmb/MiniCPM-V-4_6/raw/main/README.md recorded 2026-09-27
"The MiniCPM-o/V model weights and code are open-sourced under the Apache-2.0 license."; no datasets field in the card
Adoption
4 high confidenceHugging Face downloads over the trailing 30 days, summed across OpenBMB's MiniCPM-V checkpoints from 2.0 to 4.6, including its own quantized builds. Three speculative-decoding helper repositories the search also returns are counted too, and are negligible.
- https://huggingface.co/api/models?author=openbmb&search=MiniCPM-V&limit=100 recorded 2026-09-27
34 openbmb repositories matching MiniCPM-V with 1,238,521 downloads in the trailing 30 days; MiniCPM-V-4.6 alone 388,634
Capability
3 high confidenceMiniCPM-V understands single images, several images together and video, and its tool calling lets it hand work to other software. It documents no GUI operation and no audio input, which puts it two steps below Qwen-Omni.
- https://arxiv.org/abs/2509.18154 recorded 2026-09-27
MiniCPM-V 4.5 abstract: surpasses GPT-4o-latest on OpenCompass and "achieves state-of-the-art performance among models under 30B size" on VideoMME
- https://huggingface.co/openbmb/MiniCPM-V-4_6/raw/main/README.md recorded 2026-09-27
"It inherits the strong single-image, multi-image, and video understanding capabilities of MiniCPM-V family"; max_num_frames default 128; tool calling example; Artificial Analysis Intelligence Index 13 against 10 for Qwen3.5-0.8B
Verified 2026-09-27