FLUX 3: what Black Forest Labs shipped, and what is still just announced

BFL's pitch is a single multimodal model instead of a stack of separate ones. The same weights are meant to produce a still image, a 20-second video with its own synced audio, an edit of an existing image, and legible text inside the frame. The launch tagline reaches further still, into "world understanding" and action-prediction, framing FLUX 3 as a step toward models that understand physical space rather than just pixels.

The action-prediction angle is real but off to the side for our readers. BFL announced a companion robotics model, reported as running in Audi production facilities. We mention it only as colour: it tells you where BFL thinks this is heading, and it has nothing to do with whether you can make a picture. Do not let the robotics headline imply the image tool is ready. It is a different product on a different timeline.

The 20-second video with native audio is the new consumer-facing claim, and it is ahead of where most rivals sit on clip length plus integrated sound. We have not tested it, so we are not going to grade it. When we get access, we will run our own batch and report what breaks.

The reason FLUX matters to this audience is open weights. People run FLUX on their own machine, in ComfyUI, with no account and no telemetry, because BFL shipped downloadable models. So the only question that counts for a local user is whether FLUX 3 is one of those, and the answer right now is no.

The open-weight FLUX 3 Dev backbone is explicitly deferred to "later" with no committed date. Early access is API-and-partner only. If your workflow depends on local generation, FLUX 3 gives you nothing today, and we would not plan around it until BFL publishes weights. Announcements are not downloads.

None of that makes FLUX 3 vapourware. BFL has a strong track record of eventually opening weights. It just means the useful move today is to keep using what is already open, and check back here when the Dev release lands.

For open-weight FLUX right now, the current line is FLUX.2. FLUX.2 [dev] is a 32B image generation and editing model released on 25 November 2025, downloadable but under a non-commercial licence, so read that before you build anything paid on it. FLUX.2 [klein], released 15 January 2026, is the fast, small family: the 4B weights are Apache 2.0, the permissive open option, while the 9B weights are non-commercial. Klein runs in around four inference steps and targets sub-second, interactive use.

So the honest free-tier reality for the FLUX family is: yes, there is a real open route, but the licence is the catch. The permissive, commercial-friendly weights are the small klein 4B model, not the flagship. If you need commercial rights on the bigger models, you are on a licence, not a free download. That distinction gets glossed over constantly, and it is exactly the kind of buried ceiling worth checking before you commit.

If you want an open-weight local option outside the FLUX family entirely, our Z-Image guide covers generating locally with no account and no telemetry. For video, where FLUX 3's gated access leaves most people out for now, our free AI video generators roundup and our cost guides are the practical place to look while you wait for the FLUX 3 image and open releases to ship.

We build and run an AI companion platform, so we plan real production pipelines around these models, and "announced" versus "shipped" is a distinction that costs money when you get it wrong. A multimodal frontier launch generates a wave of coverage that quietly treats the announcement as availability. For FLUX 3, that gap is unusually wide: a big multimodal reveal, but only a gated video preview you can touch.

That is why this page leads with the availability caveat and refuses to grade output we have not run. It is more useful to tell you plainly that the image model is not out and the open weights are deferred than to publish a confident review of a demo reel. When FLUX 3 image generation opens and when the Dev weights land, we will test both and rewrite this with what we find.