New Free AI Can Watch Videos and Answer Questions About Them

A team of researchers has released Video-MOPD-8B, a free, openly downloadable AI model whose whole job is understanding video. Today's AI is good at reading text and looking at single photos, but video is harder: it moves, it has sound, and the important part is often a few seconds buried in an hour of footage. This model is built specifically for that problem.

The clever part is how they trained it. Rather than teaching one AI everything at once — which usually makes it mediocre at all of it — they trained three specialists: one for finding moments in time, one for general video comprehension, and one for science and math reasoning about what's on screen. Then they used a technique called distillation (a big AI teaching a smaller one) to merge those skills into a single model that fits on modest hardware. A filtering step called RAIS only uses training examples where the teachers were reliable and clearly better than the student, which keeps the lessons clean.

The result, according to the paper, matches or beats other open models of similar size on video tasks. That matters because "similar size" means it can run on a single powerful graphics card rather than a data center, putting serious video AI within reach of small companies, universities, and independent developers.

What it doesn't mean: this isn't a product. There's no app, no subscription, no guarantee it works on your messy home videos. Academic benchmarks are tidy compared with real footage, and the paper is a technical report, not a peer-reviewed study. But open weights mean anyone can build on it — which is usually how yesterday's research becomes next year's feature in your phone's photo app, your video editing software, or your security camera system.