Black Forest Labs has spent two years being the company everyone quietly relied on. FLUX powered a generation of image tools without ever being the loudest name in the room. On July 23, 2026, that changed. FLUX 3 is not another text-to-image checkpoint. It is a single multimodal model that generates video with sound, edits images, renders readable text, and predicts robot actions — all from one set of weights.
This is the most ambitious thing BFL has shipped, and it reframes what a "visual" model is supposed to do.
One model, four modalities
The pitch behind FLUX 3 is philosophical before it is technical. BFL argues that no single modality gives a complete description of the world. Images capture spatial structure at one instant. Video restores time and exposes physical dynamics. Audio reveals the causal link between a mechanical event and the sound it makes. Each is a lossy projection of the same underlying reality.
Train on all of them at once and the modalities start constraining each other. The sound has to match the impact. The motion has to obey the mass. FLUX 3 is the first model BFL says it built entirely on that principle — and the payoff shows up in the two things prior video models fumble most: human facial expressions and associating sounds with the physical events that cause them.
What FLUX 3 Video actually does
The headline capability is length and sound in one shot. FLUX 3 Video generates clips up to 20 seconds long in a single generation, with native audio — not a silent clip you dub later.
The supported modes cover the full production spectrum:
| Mode | What it does |
|---|---|
| Text-to-video | Prompt in, clip out |
| Image-to-video | Animate a still |
| Video-to-video | Restyle from a reference clip |
| Keyframe-to-video | Controlled transitions between frames |
| Video-audio continuation | Extend an existing clip and its soundtrack |
On top of that, BFL lists multilingual dialogue, agentic chaining of clips into multi-shot sequences, and strong typography with animated designs. In other words, it is aiming past "make me a video" and toward "assemble me a scene."
The benchmark that matters
BFL published preliminary human-preference results using 10-second, 720p text-to-video clips with audio. The numbers are lopsided at the top and honest at the bottom:
FLUX 3 was preferred over Luma Ray 3.2 in 93% of comparisons and over Runway Gen-4.5 in 77%. Against Grok Imagine Video the figure is up to 69%, then Kling v3 Pro at 60%, Happy Horse v1 at 59% and Happy Horse 1.1 at 57%. Against Seedance 2.0 and Gemini Omni Flash the result is 52% — close to a coin flip.
That last line is the credible part. A vendor claiming a 93% win over one rival and a statistical tie with two others is describing a real frontier, not a marketing chart. FLUX 3 leads, but the top of the video-generation field is now crowded and close.
Self-Flow: the method underneath
FLUX 3 is built on Self-Flow, BFL's method for aligning multimodal generation and understanding in one architecture. Self-Flow combines the flow-matching objective with a self-supervised feature-reconstruction objective. The reference implementation on GitHub is Apache-2.0, uses SiT-XL/2 with per-token timestep conditioning, trains with a 25% per-token mask ratio, and self-distills from an EMA teacher at layer 20 to a student at layer 8.
One important caveat: that public checkpoint is an ImageNet 256×256 research model, not FLUX 3. Self-Flow itself was introduced back in March 2026. What is new here is not the recipe — it is the scale. BFL says it "significantly scaled up compute and data resources" to train across video, images, and audio simultaneously.
The compute split tells you where the difficulty lives: video prediction consumes over 95% of the training compute, while audio is under 0.5% of the tokens. Sound is cheap to add once the model already understands time and motion.
The robotics surprise
The quietest claim is the most interesting. The same backbone drives FLUX-mimic, a robot control policy that runs in under 80 ms on a single RTX 5090. A generative visual model that also outputs robot actions, fast enough for real-time control on one consumer GPU, is the kind of detail that reads small today and large in a year. If a video model has genuinely learned physical dynamics, action prediction is not a separate problem — it is the same problem pointed at a motor.
The catch: you probably can't use it yet
Temper the excitement with the rollout plan, because it is deliberately staged. As VentureBeat noted, this is a limited release to start. Access is gated in a specific order:
- FLUX 3 Video and Action — available now in early access
- Image generation — early access in the following weeks
- Open-weight FLUX 3 Dev backbone — planned for release last
For a company whose reputation was built on shipping open weights, saving the open backbone for last is a notable shift. It signals both the commercial value BFL sees in this model and the safety surface area of a system that generates convincing talking-head video with matched audio.
The Bottom Line
FLUX 3 is the clearest sign yet that the neat boundaries between image, video, audio, and action models are collapsing into a single object. BFL's bet — that jointly modeling all four makes each one better — is backed by a real benchmark lead and a genuinely surprising robotics result. The gated rollout means most people will read about FLUX 3 long before they touch it. But the direction is unmistakable: the next generation of "image" models will hear, move, and act.
