🌲 FLUX 3: What's Confirmed and What Isn't Yet (2026)

Black Forest Labs shipped FLUX 3 on July 23 and gated it the same day. One network that generates images, video, audio, and robot-action predictions, with video up to 20 seconds carrying native sound — and, for now, a waitlist between you and all of it.

The launch-week coverage blurred announced and shipped into one happy paragraph, which is exactly the confusion that had people hunting for phantom Seedance 4K settings in June. Same treatment here: the ledger of what BFL has actually confirmed, what's vendor-claimed, and what doesn't exist yet.

Confirmed and usable

FLUX 3 Video, in gated early access. Text-to-video, image-to-video, video-to-video, and keyframe-to-video, up to 20 seconds per generation. Every video output carries native audio generated in the same inference pass, from the same flow-matching framework as the frames — the audio is not a bolt-on second model. Multilingual dialogue and text rendering are on the confirmed list.

Preliminary evaluations reference 720p output. Not 1080p, not 4K. 720p.

Continuation is the sleeper feature: FLUX 3 can extend existing video and audio, carry a character into new scenes, and chain clips into multi-shot sequences BFL says can run several minutes. If the character-carry holds up under real use, that's the feature working filmmakers actually needed — the 20-second cap matters less when shot two remembers shot one.

Confirmed but vendor-run

BFL published preference benchmarks for FLUX 3 Video against the field: preferred 77% of the time over Runway Gen-4.5, 93% over Luma Ray 3.2, 69% over Grok Imagine, and 60% over Kling v3 Pro.

Two readings of the same chart. The favorable one: nobody posts a 93% number against Luma unless something real is happening. The skeptical one: these are BFL's own evaluations, prompt set unpublished, and the Kling margin — 60% — is close enough to a coin flip that on Kling's home turf of motion it may be a wash. Vendor benchmarks at launch are marketing with error bars. I'll trust the independent rankings once early-access clips pile up.

Announced, not shipped

FLUX 3 Image — synthesis and editing across styles and resolutions, "early access phase in the following weeks." Notable that the image model, BFL's home category, trails the video model out the door.

FLUX-mimic — the video-action model for robotic manipulation, built with mimic robotics. Fascinating, and irrelevant to anyone making videos.

FLUX 3 Dev — the one this site cares about: open-weight access to the multimodal backbone, rolling out "over the next few weeks and months." No date, no license text, no parameter count. FLUX.1's dev weights carried a non-commercial license, so read the eventual terms before building a business on the assumption of another Wan-style Apache release. A single open-weight model that does image, video and audio together would be a first — which is exactly why the license detail will decide whether it matters.

Pricing — none published. Access runs through the early-access program and "APIs and private weight access." Check current pricing when the gate opens; anything you read before then is a guess.

By the numbers

How we verified: every row traces to the official BFL announcement or its VentureBeat launch coverage, both read this week. Nothing here comes from creator speculation, and the community numbers you'll see on X about resolution tiers and credit costs are exactly that.

My honest read: the sample reels are the most film-like output any lab has shown this year, and I still can't render a single frame with it. FLUX 3 is currently a portfolio, and portfolios don't have failure modes. The moment early access widens, we'll run our standard test prompts and report what breaks — hands, physics, or the several-minutes chaining claim, in that order of suspicion.