Summary: Black Forest Labs announced FLUX 3 on July 23, 2026 as its first multimodal foundation model, built on its Self-Flow approach and initially available through Early Access. It generates up to 20-second clips with native audio across text-to-video, image-to-video, and video-to-video modes; preliminary benchmark comparisons favor FLUX 3 over several named rivals. Its FLUX-mimic action decoder already runs production tasks on Audi lines, with the complete system reported at 101 milliseconds.
Black Forest Labs announced FLUX 3 on July 23, 2026. It is the company’s first multimodal foundation model, it sits in Early Access now, and it is the first one built from the ground up on the company’s own “Self-Flow” approach. The argument at its center runs like this: video, images, and audio are each “a projection of the same underlying reality,” so rather than train three tasks apart from one another, a single architecture ought to pick up generation and understanding together, treating every modality as evidence about one world.
The second post that landed the same day may prove the more consequential of the two over time. FLUX-mimic, a video-action model sitting on the FLUX 3 backbone, is already on an Audi production line doing real work. When a generative-media company steps into robot control, that tells you more than any benchmark number will.
Three Modalities on a Single Backbone: Betting on Self-Flow
Training for FLUX 3 drew on tens of millions of hours of general video alongside manipulation-focused footage, and video prediction by itself absorbed over 95% of the training compute. Self-Flow’s job is to align multimodal generation and understanding inside one architecture, scaled up with significantly more compute and data than the company’s earlier image-only line. The wager underneath: understanding and generation are two uses of one world model, not two products. A small detail exposes the price of that wager. Once action prediction entered the training mix, video quality fell by as much as 10% — then clawed all the way back across 3,500 training steps. One backbone, then, can carry both content creation and robot control.
Video: What It Does, and How It Scores So Far
Video carries the headline: clips running up to 20 seconds with native audio, spanning text-to-video, image-to-video (animation or reference-based), video-to-video with character carryover, keyframe transitions, multilingual dialogue, agentic multi-shot chaining, and typography solid enough for animated design work. Image synthesis and editing round out the launch, with complex prompts handled better and multilingual text rendering improved.
One caveat comes attached to the benchmark figures: they are preliminary. Human raters put 10-second 720p text-to-video clips with audio head to head. FLUX 3 came out ahead of Runway Gen-4.5 in 77% of comparisons and of Luma Ray 3.2 in 93%; the margin reached up to 69% over Grok Imagine Video and 60% over Kling v3 Pro, while against Seedance 2.0 and Gemini Omni Flash it stood at 52%.
FLUX-mimic: Turning a Physics-Aware Model Into a Robot Controller
FLUX-mimic comes out of a collaboration with the robotics company mimic robotics. A lightweight action decoder sits on the FLUX 3 backbone and reads robot actions straight out of the same world representation the model draws on when it generates video. The reasoning goes: any model capable of generating convincing video must already have grasped contact, motion, weight, and cause and effect, and that grasp can be turned toward control.
The figures: on a single RTX 5090 the backbone runs under 80 ms, and the authors say the complete system reacts in 101 ms. Even with the backbone frozen it outperforms prior vision-language-action models, needs far less demonstration data, and pulls itself out of failed grasps without ever being shown how. Audi has put it through production tasks — kitting parts into trays, assembling electronic control units, handling soft materials such as seals and cables — work the company says conventional robotics could not do. Christoph Schneider of the Audi Production Lab gave the robots credit for solving “complex soft-body manipulation work that would have been simply impossible with conventional robotics.”
How to Get It: Early Access, APIs, Open Weights
FLUX 3 Video is in Early Access now, with image Early Access expected over the weeks that follow. The launch plan spans APIs and private weights for video and image, partner access for action beginning with mimic robotics, and an open-weight “FLUX 3 Dev” backbone. For developers accustomed to running weights on their own hardware, that final item is the commitment that counts.
What This Changes for Creators and Engineering Teams
For creators, the practical change is audio-native video paired with typography and multi-shot control that actually work, which trims dubbing and storyboarding effort; dialogue in multiple languages generated along with the clip removes an entire localization pass. For engineering teams, Self-Flow and FLUX-mimic sketch out a route: rather than keep separate models for video generation and robot control, one visually grounded backbone might do both jobs. When the open-weight release arrives, and under what license, will set the pace for the whole ecosystem.