VideoLLaMA3-7B
Answers questions about a video clip or describes it in text. DAMO-NLP-SG built it on Qwen2.5-7B with a SigLIP-based vision encoder, and its training centers on image-text data.
Overview
VideoLLaMA3-7B comes from DAMO-NLP-SG, the Language Technology Lab at Alibaba DAMO Academy. You hand it a clip and a question or instruction, and it replies in text. It pairs a vision encoder tuned from SigLIP (siglip-so400m-patch14-384) with the Qwen2.5-7B language model.
The VideoLLaMA 3 paper calls its approach vision-centric. Rather than gathering huge video-text sets, the team built large, high-quality image-text datasets, and video-text data enters only in the third of four training stages. The fourth stage is fine-tuning focused on video. When a clip comes in, the model drops vision tokens that barely differ from the neighboring frame, which keeps the representation of a long video compact.
DAMO-NLP-SG reported that, as of late January 2025, VideoLLaMA3-7B was the best 7B-sized model on the VideoMME and LVBench leaderboards. Its card lists VideoLLaMA3-7B-Image as one of its base models, and Wiro runs that image checkpoint on a separate page for still photos.
What you can build
Clip libraries become searchable once each video has a written description to index. Moderation dashboards for user uploads can show a short account of what happens in a clip before a person opens it. Support and accessibility features can answer a direct question about a recording, for instance what a person in the clip is doing, or put a text summary next to the player.
Inputs
- Video (required): upload a file or give its URL.
- Prompt: the question or instruction. Defaults to "Please describe this video in detail."
- Max new tokens (advanced): the upper limit on answer length, from 10 to 4096, default 128.
Outputs
Plain text. On this page, a city stadium clip returns a short scene description: white roof, skyscrapers behind it, cloudy sky, trees around the grounds. A desert clip gets one sentence saying a man is vacuuming the sand, which the model calls unusual and impractical.
Recommended settings
128 tokens covers a short paragraph like the stadium reply. For a detailed account of a longer clip, raise the limit to a few hundred tokens so the answer does not stop mid-sentence.
Limitations
- English is the only language listed on the model card.
- The documented design reads video frames and has no audio input, so speech and other sound in the clip do not inform the answer.
- There is one video slot per request.
Safety & compliance
The model card lists the Apache 2.0 license. Before using it in a commercial product, read the license note in the project repository: it calls the service a research preview intended for non-commercial use only, subject to the Qwen model licenses, the terms of use for data generated by OpenAI and Gemini, and ShareGPT's privacy practices. If a video shows a real person, make sure you have that person's consent.
Example prompts
Great starting points for VideoLLaMA3-7B.
API quick start
Run VideoLLaMA3-7B with a single API call.
{
"prompt": "Please describe this video in detail.",
"inputVideo": "https://your-cdn.com/input.mp4",
"max_new_tokens": 128
}curl -X POST "https://api.wiro.ai/v1/Run/wiro/videollama3-7b" \
-H "Content-Type: application/json" \
-H "x-api-key: YOUR_WIRO_API_KEY" \
--data-binary @- <<'JSON'
{
"prompt": "Please describe this video in detail.",
"inputVideo": "https://your-cdn.com/input.mp4",
"max_new_tokens": 128
}
JSONOther models you might like
Explore similar video generation models.

