Black Forest Labs has launched FLUX 3, a multimodal flow model for video generation that outperforms Seedance 2.0, Gemini Omni, and Grok Imagine. The model supports text-to-video, image-to-video, and video-to-video generation with native audio. This release marks a significant advancement in AI video generation.
SynthePulse Insight · AI deep reading
Version 2 · 4 sources
Black Forest Labs releases Flux 3, the first model to jointly generate video and native audio, but internal tests show limited performance advantage; the open-source plan is the biggest highlight.
Black Forest Labs released Flux 3 on July 23, 2026, positioning it not merely as a text-to-video or text-to-image model, but as a 'real-world model.' The team places images, video, and audio into the same multimodal flow architecture, attempting to learn 'how the world works.'
According to BFL, each modality is considered a 'projection' of the same reality: images provide spatial structure, video provides time and dynamics, audio provides causal cues, and language provides abstract instructions. During joint training, mutual constraints between modalities (e.g., sound must match impact) are closer to 'world representation.'
This architecture is based on the Self-Flow method, using a multimodal Transformer. BFL states that the same base can be used simultaneously for content generation and embodied/action prediction.
Flux 3 is the first to support generating video with native audio, up to 20 seconds. In BFL's internal tests, for 10-second 720p video preference rate, Flux 3 leads most competitors, but the advantage is limited.
Specifically, preference rate relative to Luma Ray 3.2 is 93%, to Runway Gen-4.5 is 77%, to Grok Imagine Video is 69%, to Kling v3 Pro is 60%, to Seedance 2.0 is 52%, and to Gemini Omni Flash is 52%.
Note that these performance figures are from BFL internal tests and have not been independently verified. Independent tests have not yet been conducted, so actual performance may be uncertain.
BFL plans to release open-source weights version Flux 3 Dev, currently only available via early access application. According to reports, API access will be opened in the coming weeks, followed by the open-source version.
The open-source timeline may change, but this plan has already attracted widespread community attention.
BFL collaborated with mimic robotics to develop Flux-mimic, a video action model based on the Flux 3 backbone for general dexterous manipulation. The model has been tested in Audi production tasks and can be deployed on a single local GPU.
This demonstrates Flux 3's potential to extend from content generation to embodied intelligence.
This article primarily relies on BFL official release information and media reports. Performance data are from BFL internal tests and have not been independently verified. The open-source plan timeline may change.
Flux 3 achieves breakthroughs in multimodal unified learning and native audio-video generation, but performance advantage is limited; the open-source plan is the biggest highlight.
Primary report
Primary source
Also covered by 3 sources