Many video models can describe frames, but understanding what happened often depends on connecting sound to the right moment and speaker. This work makes joint audio-visual reasoning its central test, rather than treating video as a sequence of images with optional transcripts. Its benchmark results are promising, though they remain claims from the model's authors until independently reproduced.
Key Findings
- The 2B, 4B, 9B, and 35B-A3B variants share a bilingual, multimodal design and support both direct responses and explicit step-by-step reasoning. The range makes it possible to compare capability across model sizes.
- Training uses approximately 1.2 trillion multimodal tokens and a staged post-training pipeline. The scale signals a substantial training effort, but the abstract does not isolate which training component contributes most to performance.
- Across more than 120 benchmarks, the authors report strong audio-video understanding and competitive results on audio, silent video, images, and text. They claim the 35B-A3B model sets a new open-source high on omni-modal video tasks, while the 9B version surpasses Qwen3-Omni-30B-A3B-Instruct in their evaluations.
- The paper also releases evaluation and reproduction materials, an important step for testing results across different video sampling, audio, and reasoning conditions.
Who Should Read It
A good fit for researchers comparing open multimodal models, especially on tasks where audio and video must be reasoned about together. Treat the reported rankings as a starting point, not a settled leaderboard: the available results are author-reported, and independent replication is still needed. Teams seeking established production performance or independently verified deployment costs should wait for broader external testing.