Video understanding is kind of new, especially if done well, and hopefully working well with UI and UX, that'd be great. Current agents already struggle a bit with 2D space with normal screenshots of unconventional UIs, wonder if this model would do better with actual recordings of navigating and using applications, feels like it could help a bunch with understanding UX at least hopefully. Will be fun to play around with :)
Sure, but again, it's a micro 3B model. Perhaps it can't be used for general video work, but it might be able to do basic edits like remove an object from a table in a shot.
Not that surprising if the reason you have virtually unlimited compute and programming resources is that you work at the leading short form video app company. They could also have chosen nót to open source it.
Do you find the video understanding work there also to be 'silly little slop', or did you only look at the gifs on the page and not read about the understanding work in a 3B model?
This is not ground-breaking by any means, but achieving this in a 3B model and sharing the approach + weights advances engineering and certainly more contribution that 'silly little slop videos' imo.
Video understanding is kind of new, especially if done well, and hopefully working well with UI and UX, that'd be great. Current agents already struggle a bit with 2D space with normal screenshots of unconventional UIs, wonder if this model would do better with actual recordings of navigating and using applications, feels like it could help a bunch with understanding UX at least hopefully. Will be fun to play around with :)
Seems like the video output is crippled. Resolution is low (720 or so), as is the frame rate. The samples are shown up-scaled and frame-interpolated.
Why do that? Seems strange to be building sub-hd resolution video models in 2026.
Sure, but again, it's a micro 3B model. Perhaps it can't be used for general video work, but it might be able to do basic edits like remove an object from a table in a shot.
It’s not a micro model at all, it requires 40gb of VRAM. The 3B is just the active parameters.
What’s SOTA for video understanding? AFAIK most video search is powered by transcription and not the actual video. This seems impressive.
Great quality, forked and going to try
Any plans to port to sglang or vLLM?
vllm-omni support is on the way : )
Nice work. Wish they had picked another name given how popular lance/lancedb is.
[flagged]
[flagged]
Not that surprising if the reason you have virtually unlimited compute and programming resources is that you work at the leading short form video app company. They could also have chosen nót to open source it.
Do you find the video understanding work there also to be 'silly little slop', or did you only look at the gifs on the page and not read about the understanding work in a 3B model?
This is not ground-breaking by any means, but achieving this in a 3B model and sharing the approach + weights advances engineering and certainly more contribution that 'silly little slop videos' imo.
It’s not a 3B model, it has 3B active parameters. The full model is much larger.
That's true, I should have mentioned active. Actual params are closer to 12B-14B likely, given the 40GB VRAM usage.
last dance for lance vance!
:D