Black Forest Labs made its name with FLUX.1 and FLUX.2 — image models that became the default open-weight choice for a large slice of the industry, powering features inside Adobe Photoshop, Picsart, and countless startups. On July 23, the Freiburg lab announced something stranger and more ambitious: FLUX 3, a single model jointly trained across images, video, audio, and action prediction, aimed at what the company calls “visual intelligence.”
It is the most aggressive unification bet in AI right now. It is also, for the moment, almost impossible to try.
One model, four modalities
FLUX 3’s thesis is simple to state and hard to execute: images, video, audio, and actions are partial projections of the same underlying physical reality. Train on all of them jointly, in one architecture, and each modality teaches the others — the sound has to match the impact, the motion has to obey the mass, the future has to follow from the past.
Technically, FLUX 3 is a diffusion transformer (DiT) trained with what BFL calls Self-Flow, a self-supervised flow-matching framework, across tens of millions of hours of general video, hundreds of thousands of hours of manipulation video, plus images and audio. The shipping modality is video: up to 20 seconds with synchronized native audio. Image generation is promised “soon.” And the action-prediction head — the part that extends the same backbone toward robotic vision and action — is the long-term play.
- Video: up to 20 seconds, 720p in early evaluations, with synchronized audio generated natively rather than dubbed afterward.
- Image: synthesis and editing through the same backbone — arriving after video.
- Audio: learned jointly; generated in sync with video.
- Action prediction: the same architecture extended to predict robot actions from visual input.
Why unification, why now
The industry has spent years assembling multimodal systems from parts: a language model here, a vision encoder there, a vocoder for audio, a separate policy for robotics. It works, but the seams show — lip-sync drift, physics violations, the general sense that the model does not understand the scene so much as translate between descriptions of it.
BFL’s argument is that the seams are the problem, and the fix is joint training from the start. If a model learns video and audio together, it cannot help but learn that a slammed door looks and sounds a particular way — the cross-modal constraints teach physics. Whether that philosophical elegance translates into better outputs is the empirical question FLUX 3 exists to answer.
The access problem
Here is the catch: you almost certainly cannot use FLUX 3 yet. Initial access runs through a closed-door Early Access program — anyone can request access, but BFL approves each applicant individually. There is no public API, no partner distribution, and no general availability date. FLUX 3 Video and FLUX 3 Action are the gated modalities; FLUX 3 Image follows in the coming weeks; a FLUX 3 Dev open-weight backbone is promised “later.”
This is a sharp break from BFL’s history. FLUX.1 and FLUX.2 won their position precisely through open access — weights you could download, fine-tune, and build businesses on. FLUX 3’s gated launch suggests the lab believes this generation is different in kind: more capable, more dual-use, and more valuable to meter. The per-second API pricing (when it opens) points the same way — BFL is positioning FLUX 3 as premium infrastructure, not a community download.
The robotics shadow
The action-prediction head is the part competitors will study most closely. “Visual intelligence,” in BFL’s telling, means models that perceive, predict, and act across physical and digital environments — and action prediction is the bridge from generative media to physical AI. Hundreds of thousands of hours of manipulation video in the training mix is not there for prettier clips; it is there to teach the model how hands move objects, which is the raw material of robotics.
The timing aligns with the industry’s broader pivot. 2026 has seen the conversation shift from language models toward world models, simulation, and physical AI — adjacent fields that BFL explicitly frames as “interconnected expressions of the same underlying capability.” FLUX 3 is the lab’s attempt to own that convergence rather than watch it from the image-generation sidelines.
What is real and what is promised
Discipline is warranted. What is real today: the announcement, the architecture description, the training-data scale claims, the early-access program, and the per-second pricing model. What is promised: the image modality, general availability, the open-weight Dev backbone, and the robotics payoff. The parameter count was not disclosed at announcement, and license terms for FLUX 3 had not been published — both worth watching, given how much of BFL’s reputation rests on openness.
Early community analysis has focused on the pricing math: per-second billing at multiple resolutions, sold only through BFL’s own API and licensing tiers. For studios and developers, the question is whether unified generation is worth a premium over the current stack of specialized models — Veo or Kling for video, a separate TTS engine for audio, FLUX.2 for images. Unification has to earn its keep in output quality, not just architectural elegance.
The bet, stated plainly
FLUX 3 is a bet that the future of generative AI is not a toolbox of specialized models but a single system that understands the world well enough to render it in any modality — and eventually act in it. If BFL is right, the lab that owned open image generation is positioning to own the next layer: the foundation for both synthetic media and physical AI.
If the bet is wrong — if joint training turns out to be an elegant idea that specialized models keep beating in practice — FLUX 3 will be remembered as the moment the lab overreached. The evidence will come when the gates open and independent evaluations replace launch materials. Until then, it is the most interesting announcement of the summer: not a bigger model, but a different theory of what models should be. For the image-model establishment FLUX 3 is challenging, see GPT Image 2.5; for the video-length race it joins, see Kling 4.0.
Featured image: VulcanSphere via Wikimedia Commons (CC0 1.0)
