Video leaderboards' current 7B champion is smaller than you think

What it does VideoLLaMA3 is a collection of multimodal models—2B and 7B parameter variants built on Qwen2.5—that accept images or video clips and answer questions about them. You feed it a video path with a target frame rate and maximum frame count, or pass in images and text prompts, and it generates responses via standard Transformers inference. The project also provides image-only checkpoints and a separately downloadable tuned SigLIP vision encoder.

The interesting bit As of late January 2025, the 7B model ranks first among 7B-sized entries on both the LVBench and VideoMME leaderboards, which test long-video and multi-modal reasoning. The team also released the synthetic re-captioning dataset VL3-Syn7M used to train it, so you can inspect the training diet rather than guessing.

Key highlights

• 2B and 7B parameter models with dedicated image-only variants
• #1 ranked 7B model on LVBench and VideoMME leaderboards (as of Jan 2025)
• Handles fine-grained tasks: visual referring, grounding, chart analysis, table understanding, and document recognition
• Tuned vision encoder (SigLIP-NaViT) released separately for reuse in other pipelines
• Runs via Hugging Face Transformers with Flash Attention 2 and bfloat16

Caveats

• Requires trust_remote_code=True and a fragile dependency stack (PyTorch 2.4.0, flash-attn 2.7.3 for CUDA 11.8)
• README is strong on benchmark screenshots but sparse on architecture and training details
• Leaderboard claims are bold for a relatively new repo; verify against your own video tasks

Verdict Try it if you need a compact, video-capable VLM that also handles documents and charts. Look elsewhere if you want a fully transparent training recipe or an architecture without custom remote code.