starzero.ai retweeted
Gemini 4 Argon just took #1 on our video understanding benchmark. For most of this year, #1 was a model I’ve barely seen mentioned online: ByteDance's Seed 2.0 Lite.
Some context. We’ve had an internal video understanding benchmark for a long time. It changed a lot, but the goal never did: tell us whether a new model is good enough for us to adopt. This year the best models went from 50 to 81 on it, so it’s time to build a harder one. Before we do, we’re sharing the results. If there’s enough interest, the next one might be public, open source, or a mix. We’ll see.
Testing every model was a blast. We watched Gemini take #1, lose it, and take it back. We saw great LLMs that were never promoted for video do really well. And we saw some companies that market themselves on video understanding score below most of the open-source models.
Credit where it’s due:
- Google, for the best video understanding model and for pushing the frontier. They also have the best native video support: far fewer tokens than feeding frames, without hurting understanding. Sorry Meta, you need to work more on that.
- Meta, for catching up to the top so fast. Still very expensive for video, though.
- OpenBMB, for always pushing what a good, affordable video model can be with the MiniCPM series (congrats on MiniCPM-V-4.7).
- Liquid AI, for really efficient models. If they focused on video, they’d probably nail that too.
- Alibaba Group's Qwen, the best open-source video model, no questions asked.
- Anthropic for making such good models that can drive video editing without having native "video" skills.
- And last but not least, everyone who releases open-source models!
What you see is the tip of the iceberg. We’ve tested video mode, image mode, different frame rates, collage mode and every thinking level, to see what works and how much each one helps or hurts. If there’s interest, we’ll share more.
Which model should we test next? Request it on the page, or tell me in the comments.
And if you’re wondering why we do this: at @starzeroAI, we're building AI video editing that just works, for everyone.
starzero.ai retweeted
My Euro-fied brain just couldn't process the Fahrenheit numbers thrown at it in @Casey's YT video (link below) so I instructed @starzeroAI to make an edit with just the temperature display scenes and show them in Celsius.
Coming soon to starzero.ai. Shoot your scene, prompt the style.
Footage courtesy of youtube.com/ManCarryingThing
To help people keep up with the ("$200k LEGO Theft") Bricks & Minifigs saga we started putting all the videos we can find about it into our indexer. Catch up on @Medium : starzero.medium.com/the-bric…
We've indexed the original Reckless Ben series plus @MoistCr1TiKaL's, @Asmongold's, @LegalEagle and other takes. If you have a video (or PDF or whatever) for us to index and add to our Reckless Ben - BAM Dashboard, let us know!


