Open weights. Build with LLaVA-OneVision-2.5🥳🥳
Can a video model learn from the codec instead of from sampled frames — and actually win?
It took three steps. OneVision-Encoder showed that a vision encoder can be aligned with codec signals. LLaVA-OneVision-2 gave the first evidence inside a full VLM. Now LLaVA-OneVision-2.5, led by @GlintLabAI and @lmmslab, is where codec input finally beats frame sampling on general video tasks.
Instead of looking at 1 frame per second, the model uses the codec's bit-cost to find the regions that carry new information and packs them into a few canvases. The same 4B weights run in two modes: offline, for images and long videos, and online, where frames stream in and the model decides at every step whether to speak or stay silent.
Result: across 17 video benchmarks, codec input beats the model's own frame sampling on 13 and Qwen3.5-4B on 16. For streaming, a full-frame-rate design unique to this model lets the 4B LLaVA-OneVision-2.5 stand alongside frontier 8B–11B streaming models.
Model: huggingface.co/lmms-lab-enco…
Code: github.com/EvolvingLMMs-Lab/…
2
23
LLaVA-OneVision-2.5 is out! Co-developed and maintained by Glint Lab and @lmmslab
One model for full-frame-rate streaming and offline video analysis.
For the first time, we see clear signals that codec beats uniform frame sampling on general video tasks.
Today we're releasing LLaVA-OneVision-2.5, with major efforts from @GlintLabAI.
In OV-2, we aim to prove that codec can work for video understanding. We trained on video using both ways, frame sampling and codec, and each did better on different kinds of tasks. We observed codec can work but mostly good at temporal grounding tasks, not on all video tasks.
But we think it can work in a more unified way. Even a single image can be stored in a compressed view, like JPEG. Video is even less a pile of independent images.
In OV-2.5, we show that codec input comes out ahead of frame sampling on most of the general video benchmarks.
More deltas on OV-2.5:
1/ One model, two modes. The same model and the same chat template serve both offline and streaming video. In offline mode,
2/ In offline mode, the model is good at understanding extremely long video. 9× spatial compression and a 262k context let it read 1,000+ frames in a single pass. And it beats Qwen3.5-4B on 16 of 17 video benchmarks.
3/ In streaming mode, the model is proactive and knows when to speak. It watches video streams at 24 fps, and the 4B model ties an 11B streaming model on OVO-Bench and more than doubles the next best score on StreamArena proactive interaction.
Weights and training code are open today, and the technical report is coming soon.
Github: github.com/EvolvingLMMs-Lab/…
HF: huggingface.co/lmms-lab-enco…
Most video models sample frames. LLaVA-OneVision-2.5 watches every frame — on-device. Codec sends only what changed; fork-and-commit keeps one model watching while it answers. 2B/4B · RK3588+RK1828 · 0.70/0.82s to judgment · 1080p60 #VideoLLM huggingface.co/lmms-lab-enco…


