Skip to content
← Models
Open weightsDAMO Academy (Alibaba)/10 trustVerified

VideoLLaMA3

Frontier multimodal foundation model for image and video understanding, built on SigLIP and Qwen2.5.

Visit official page →

About

VideoLLaMA3 is DAMO Academy's video-understanding foundation model, built on SigLIP for vision and Qwen2.5 for language, with design lessons borrowed from LLaVA-OneVision, InternVL2 and Qwen2-VL. The January 2025 technical report positions it as a general image-and-video model with spatial-temporal modeling, continuing the Video-LLaMA line that started in 2023.

The code is Apache 2.0, but the weights are released as a research preview for non-commercial use, so commercial teams should read the license carefully. VideoLLaMA3-7B is on Hugging Face for self-hosting. Published evaluations are author-reported only, with no independent benchmark numbers to cite, and the fast-moving open video-model field has since produced strong alternatives. It is still a solid reference implementation for video-language modeling, especially for researchers already in the DAMO and Qwen ecosystem.

Specs

Context window

32,768 tokens (Qwen2.5-7B base)

Pricing

Free / open weights (research preview, non-commercial per repo)

Free quota

Weights free for non-commercial research use

API access

Hugging Face (DAMO-NLP-SG/VideoLLaMA3-7B); self-hosted

Weights

Open weights

Released

Not disclosed

Benchmarks

  • Author-reported evaluations in the technical report (arXiv 2501.13106); specific numbers not independently verified

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

video-understandingmultimodalimage-understanding

Trust score — 6/10

/10 trustVerified

Scored 6/10 (TIER metrics, metrics-v1). Evidence: arXiv 2501.13106 (Jan 2025) from DAMO-NLP-SG; GitHub DAMO-NLP-SG/VideoLLaMA3; HF DAMO-NLP-SG/VideoLLaMA3-7B. Strengths: frontier-positioned open video model, builds on proven SigLIP+Qwen2.5 stack, documented lineage from LLaVA-OneVision/InternVL2. Docked for: (1) evaluations are author-reported only, no independent numbers; (2) weights are non-commercial research preview, not truly open; (3) crowded field with newer alternatives. License: Apache 2.0 (code); research-preview non-commercial terms (weights).

How the trust score is built

Full methodology →
Adoption · stars/downloads, log-scaled vs peers
5.0/10
Maintenance · recency, releases, issue load
5.0/10
Openness · license + docs
6.0/10
Manual review · sanity read + red-flag check
6.0/10
  • · Technical report arXiv 2501.13106 (2025-01-22) from DAMO-NLP-SG; SigLIP + Qwen2.5
  • · GitHub DAMO-NLP-SG/VideoLLaMA3 verified live (curl 200, Oct 2026); HF DAMO-NLP-SG/VideoLLaMA3-7B
  • · Code Apache 2.0, but weights are a research preview for non-commercial use only (repo license section)
  • · Evaluations author-reported only; no independent benchmark numbers to cite

Sources