Flux 3 previews as an open-weights video model with 720p samples

• FLUX 3 is live in a 48-hour Hermes Agent preview, where BFL says Hermes can chain shots into full short films from one prompt, according to BFL's preview post.
• The open-weight path is staged: Black Forest Labs' launch post separates FLUX 3 Video from future FLUX 3 Dev access, and venturetwins pointed followers to BFL for open-source timing.
• Early demos are already creator bait: the cleanest examples so far are venturetwins's GPU ASMR post and venturetwins's “Contaminated” trailer.
• Output quality is the launch caveat: CuriousRefuge's comparison ranked FLUX 3 behind Seedance 2.0 and LTX 2.3, and noted a current 720p limit.
• Rival models are still setting the visible bar: Artedeingenio's Seedance 2.0 clip challenged FLUX 3 on choreography, while BLVCKLIGHTai said Hailuo H3 was better with vague prompts so far.

Black Forest Labs' launch post says FLUX 3 can generate video with native audio up to 20 seconds, then splits access into Video APIs/private weights and a future open-weight FLUX 3 Dev. RuntimeWire reports the Hermes window is free for paid Nous Portal subscribers for 48 hours, with a short-film contest ending August 1. BFL also tied the preview to its San Francisco open-studio page. The funniest first use case was venturetwins's GPU ASMR test.

Access opened through Hermes Agent, not as a general model-card download. BFL's preview post says the preview runs for 48 hours and frames Hermes as a one-prompt short-film chain.

Venturetwins, who posted several early FLUX 3 clips, told one commenter it was “free right now on Hermes” in a reply. RuntimeWire narrowed that to free access for paid Nous Portal subscribers during the 48-hour window.

When asked about the workflow behind an AI horror trailer, venturetwins answered with the whole selling point: one prompt in a workflow reply.

Creators heard “open weights” because BFL used that language for FLUX 3 Dev. The official launch plan separates four lines:

• FLUX 3 Video: video and audio generation and editing through APIs and private weight access.
• FLUX-mimic and FLUX 3 Action: action prediction through selected research and commercial partners.
• FLUX 3 Image: image synthesis and editing through APIs and private weight access.
• FLUX 3 Dev: open-weight access to a multimodal backbone for content creation and action prediction.

That split is the most important launch caveat for local creators. venturetwins told followers to watch BFL for timing after saying the company had announced open source.

BFL describes FLUX 3 as a unified image, video, audio, and action model, not a narrow text-to-video generator. Its launch post lists these Video capabilities:

• Text-to-video generation.
• Image-to-video animation or visual references.
• Video-to-video generation from a reference clip.
• Video and audio continuation from input video and audio.
• Keyframe-to-video transitions.
• Multilingual dialogue.
• Multiple styles and aspect ratios.
• Agentic chaining into longer multi-shot sequences.
• Native audio on all listed video outputs.
• Typography generation and animated designs.

BFL says its preliminary Video analysis used 10-second text-to-video clips in 720p with audio, while the claimed single-generation ceiling is 20 seconds.

The first creator examples lean into texture, mood, and genre trailer grammar: glowing hardware surfaces in one clip, decayed cosmic-horror images in another.

Another early-access watcher, ozansihay, said the first creator samples looked strong enough to make FLUX 3 a possible Seedance 2.0 rival.

CuriousRefuge ran FLUX 3 against Seedance 2.0 and LTX 2.3, then put the early model behind both across dialogue scenes, product commercials, and multimodal workflows combining video, audio, and text.

The same post said FLUX 3 is currently limited to 720p. That sits awkwardly next to BFL's internal preference claim: the official launch post says FLUX 3 beat Seedance 2.0 in 52% of early comparisons, which is close enough to a coin flip for creators to keep testing their own prompts.

Artedeingenio's Seedance 2.0 sword-fight clip became the visible bar for fast motion and camera tracking, with the challenge stated directly: can FLUX 3 make something like this?

FLUX 3 was not the only model creators were stress-testing this week. After downloading FLUX 3 through Hermes, chrisfirst said he was testing references using Hailuo-03.

BLVCKLIGHTai said Hailuo H3 had done better than FLUX 3 with vague prompting so far, then added that it was too early to pick a winner.