Article: Flux 3, Black Forest Labs’ newest video-generation model, creates 20-second clips with built-in audio and, according to the lab’s own benchmark, beats every major competitor it was pitted against. The internal test gave Flux 3 a 93 % preference over Luma Ray 3.2 and a 52 % edge even against Seedance 2.0, the model it was designed to surpass.

Why the numbers matter

BFL’s benchmark, dubbed BFL Tests, pits models side-by-side on a set of prompts and asks human raters which output looks better. Flux 3 topped the list across the board:

  • 93 % preferred over Luma Ray 3.2
  • 77 % over Runway Gen-4.5
  • 69 % over Grok Imagine Video
  • 60 % over Kling v3 Pro
  • 52 % over Seedance 2.0 and Gemini Omni Flash

If the results hold up, developers and content creators could gravitate toward a single model that handles text-to-video, image-to-video and video-to-video in one package, cutting the need to stitch together multiple tools.

The tech behind the claim

Flux 3 introduces “Self-Flow,” a training method that ingests images, video frames and audio tracks simultaneously. It also claims to handle complex prompts that mix languages, allowing multilingual dialogue within a single clip. Clip chaining lets users link several 20-second segments into longer narratives, a feature BFL says addresses a common limitation of earlier generators.

Beyond media creation

BFL isn’t stopping at pure content generation. A partnership with Mimic Robotics has produced “Flux-mimic,” a version of the model that can suggest actions for physical robots. Audi is testing this video-action variant for production-line tasks, hinting at a future where generative AI informs real-world operations, not just entertainment.

The missing piece

All the hype rests on BFL’s internal testing. No third-party benchmark has been published, and the methodology behind the preference scores is not public. Until external labs reproduce the results, the claim that Flux 3 is the clear leader remains provisional.

What to watch

  • Independent evaluations – Labs or academic groups will likely run their own comparisons; any divergence from BFL’s figures will shape market perception.
  • Early-access rollout – BFL plans to release an image-generation version of Flux 3 in the coming weeks, expanding its ecosystem.
  • Audi’s production results – Concrete outcomes from the Flux-mimic trial could demonstrate whether generative video can reliably guide robotic actions.