One model, many tracks
Video with synchronised stereo audio generated as one output, not two pipelines. And a perception encoder that is 6% of the parameters and sets the ceiling on everything.
Sync is the hard part
Footsteps landing on the frame the foot lands. Speech matching mouth movement. Generating audio and video as one output rather than dubbing one onto the other is the harder approach and the correct one, and it is the specification worth noticing in that release.
21:9, 16:9, 4:3 and 1:1 is a product decision, not a research one. Cinema, television, archive, social.
Video has been the most closed of the major modalities because compute and liability are both high. Weights in the open means fine-tuning, local deployment, and no ability to withdraw it later. That is the point and the argument simultaneously.
Now the part nobody reads
Whatever the encoder fails to represent is not recoverable. Fine text, small objects, spatial relationships lost at encoding are gone, and no amount of reasoning capacity reconstructs them. The transformer is reasoning over a summary it did not produce and cannot inspect.
This explains the failure everyone has hit: a model that discusses an image fluently and gets one specific detail wrong. That is almost never a reasoning failure. It is the encoder having discarded it.
Why the reporting hides it
Multimodal benchmark scores are attributed to the model as a whole, so an encoder limitation reads as a reasoning limitation. Parameter counts and context windows get the coverage; encoder resolution decides what the system can perceive at all.
What this asks of you
Test perception separately from reasoning. Ask it to transcribe and locate before you ask it to interpret. If it cannot read the label, its analysis of the label is fiction — and that gets more consequential as models take more modalities at once.
AIBase — Visual Large Models Receive a Major Open-Source Announcement → · BentoML — Multimodal AI: A Guide to Open-Source Vision Language Models →