Polaris and Vega: Mystery AI Video Models Spark Meta Speculation
Two unnamed AI video models, Polaris and Vega, surfaced on Artificial Analysis. Here's what their outputs suggest and who might be behind them.
What are Polaris and Vega?
Polaris and Vega are two unreleased, unbranded AI video generation models that appeared on the Artificial Analysis Arena, a leaderboard where new models are often tested anonymously before their official launch. Neither company has confirmed ownership of either model. The outputs currently limited to text-to-video clips run about 10 seconds long, include audio, and render at 1080p. Based on visible differences in output style, quality, and behavior, Polaris and Vega appear to be two distinct model families, not variants of the same system.
TL;DR
- Two anonymous video models, code-named Polaris and Vega, showed up on the Artificial Analysis Arena leaderboard without any official confirmation of who built them.
- Meta is the leading guess given recent activity from its superintelligence lab, including the Muse Image model and an already-announced video model in development.
- Polaris shows strong multi-shot consistency, holding character details and scene continuity across cuts, along with solid prompt adherence for physical effects like breath fogging glass.
- Vega demonstrates multilingual dialogue and lip-sync, generating accurate Italian speech with synced mouth movements across multiple mirrored reflections in one test clip.
- Both models still show text rendering quirks, including on-screen text that flickers or pops in in one Vega sample despite otherwise clean signage and spelling.
- Kling 4 and a new VO model remain other candidates, since both are widely anticipated releases that haven’t yet gone public.
- A separate resource, Stills Lab, offers a searchable database of film stills with color palettes, lens data, and camera info useful for prompting image and video models.
Other agents ship a demo. Remy ships an app.
Real backend. Real database. Real auth. Real plumbing. Remy has it all.
Why is Meta the top suspect?
The speculation around Meta stems from a string of recent releases tied to its superintelligence lab, led by Alexander Wang. That team recently shipped Muse Image, an image generation model, and Muse Glimmer, a 30-billion-parameter open-source agentic language model. Meta also said it plans to open-source its frontier model, referred to as Muse Spark 1.2. Alongside the Muse Image release, Meta signaled that a video model was coming next. Polaris and Vega appearing on a public leaderboard shortly after fits that timeline, though nothing in Artificial Analysis’s listings explicitly ties either model to Meta.
The other major unreleased video model people are waiting on is Kling 4, from Kuaishou. Based on the style and behavior of the outputs, the creator reviewing these clips felt Polaris and Vega didn’t match what’s expected from Kling. That leaves open the possibility that one of the two is a new Google VO model, or something from a company that hasn’t been on anyone’s radar yet.
How good is Polaris in testing?
Polaris outputs reviewed so far show a level of physical and narrative consistency that stands out. In one clip, a character writes down a sequence of digits across multiple shots, and the numbers stay consistent from one cut to the next, a detail that trips up a lot of video models. In another, a dog breathes on a window and the prompt’s specific request for fogged glass and a smear mark gets rendered accurately, though the dog’s nose doesn’t press against the glass as fully as it should.
A separate Polaris test involving a banner catching wind and nearly lifting a person off the ground handled the physics correctly, keeping the person grounded rather than producing the kind of uncontrolled floating or flying artifacts common in older video models. A rock climbing clip also held up well across multiple shots of the same location, with only a minor glitch where a hand placement disappears during a reach. Polaris also handled a stylized test involving a Batman-inspired character, rendering the visual identity clearly rather than producing a vague, legally-safer variation, which suggests looser guardrails around copyrighted imagery than some competing models.
What makes Vega different?
Vega’s standout feature in testing was multilingual dialogue paired with lip-sync accuracy. In one clip, a character delivers lines in Italian, and the audio appeared to match the language correctly while lip movements stayed in sync. That same test involved a hall of mirrors with multiple reflections, all of which moved and stayed lip-synced to the speaker, a technically demanding combination of physics and audio-visual alignment.
Text rendering was a mixed result for Vega. A clip showing a sign being painted on a storefront window displayed clean, consistent lettering and design, but suffered from the text appearing to pop up rather than being painted in a physically continuous motion. When Vega and Polaris were given the same prompt (a nervous pilot reciting numbers under pressure), the Polaris version was judged to be the stronger output overall.
- ✕a coding agent
- ✕no-code
- ✕vibe coding
- ✕a faster Cursor
The one that tells the coding agents what to build.
Is this actually Meta’s next video model?
There’s no confirmation either way. What can be said is that the timing lines up with Meta’s recent release cadence, and the company had already signaled a video model was in development following Muse Image. Beyond that, attributing Polaris or Vega to any specific company is speculation based on style, quality, and process of elimination against known upcoming releases like Kling 4. Companies sometimes test unreleased models anonymously on public leaderboards specifically to gather feedback before revealing branding, so it’s also possible neither model turns out to be from a company currently on people’s shortlist.
What else is worth knowing for AI filmmakers right now?
Separate from the mystery models, a tool called Stills Lab was highlighted as a resource for people building AI-generated video and film content. It functions as a searchable library of still frames pulled from movies, TV shows, and music videos, similar in concept to existing tools like ShotDeck. Each still comes with metadata including color palette hex codes, lens information, and camera details, all useful inputs for prompting image and video generation models.
Its distinguishing feature is a visual search function: upload an image, and it returns stills from other films and shows with similar tone, color grading, or composition. That can be useful for finding stylistic references or generating prompt language grounded in the technical specifics (lighting, lens choice, color grading) of professionally shot footage rather than guessing at descriptive terms. The service offers a free tier as well as a low-cost paid plan.
Frequently Asked Questions
What is Artificial Analysis Arena?
It’s a platform where AI models, including video generation systems, get tested and compared, sometimes anonymously before an official public release. New or unbranded models occasionally show up there ahead of formal announcements.
Are Polaris and Vega confirmed to be from Meta?
No. There’s no official confirmation of who built either model. Meta is considered a likely candidate due to its recent pattern of releases (Muse Image, Muse Glimmer) and its stated plans for a video model, but this is speculation, not confirmed fact.
How long are the sample videos from Polaris and Vega?
Based on what’s appeared on the leaderboard, both models are currently limited to about 10-second text-to-video clips, rendered at 1080p with audio included.
What other companies might be behind these models?
Kling 4 (from Kuaishou) and an updated Google VO model are the other most anticipated unreleased video models. It’s also possible one or both come from a company not currently expected to release a new video model.
What is Stills Lab used for?
It’s a database of film and TV still frames with attached metadata like color palettes and camera/lens details, plus a visual search feature for finding similarly styled shots. It’s aimed at people prompting AI image and video tools who want grounded stylistic references.