FLUX 3 announced in Early Access

25 juillet 2026 · Kris Wilhem

FLUX 3 announced: image, video, audio and action in a single model

Black Forest Labs has just announced FLUX 3, a new generation of multimodal foundation model.

Video is already entering Early Access, image will follow in the coming weeks, and a FLUX 3 Dev open-weights version is officially planned.

One model for several media

With FLUX 3, BFL is no longer simply presenting an image generator.

The model is trained jointly on:

  • image
  • video
  • audio
  • language
  • action prediction

The aim is to build a more coherent representation of the world: understanding objects, their movements, the sounds they produce and the way a scene evolves over time.

An image provides the visual structure. Video adds movement and continuity. Audio links visible events to their sound environment. Language makes it possible to direct the whole.

FLUX 3 Video enters Early Access

The first accessible version is FLUX 3 Video.

BFL announces in particular:

  • up to 20 seconds of video
  • text-to-video generation
  • image-to-video
  • video-to-video transformation
  • clip continuation
  • use of keyframes
  • character consistency across several scenes
  • generation of native synchronised audio
  • multilingual dialogue
  • motion design and animated typography

Access remains for now reserved to a selection of testers.

FLUX 3 Image will follow

FLUX 3 Image is due to arrive next with generation and editing functions.

BFL highlights a better understanding of complex prompts, more reliable text generation and improved handling of multilingual content.

For the moment, no precise information has yet been given on:

  • model size
  • VRAM consumption
  • local performance
  • ComfyUI integration
  • Diffusers compatibility
  • fine-tuning
  • LoRAs

FLUX 3 Dev is confirmed

This is probably the most important announcement for local users.

Black Forest Labs confirms the arrival of FLUX 3 Dev, presented as an open-weights multimodal backbone.

It would therefore not be just a new text-to-image checkpoint, but potentially the open foundation capable of handling several modalities: image, video, audio and action.

The release date, the licence, the hardware requirements and the fine-tuning possibilities remain unknown.

How Axiris reads it

For Axiris, the real interest of FLUX 3 does not lie only in an improvement in visual quality.

The major change is the possibility of linking, within a single system, several stages that are separate today:

designing an image, modifying it, preserving a visual identity, animating a scene, producing its sound environment and maintaining continuity across several shots.

This approach directly matches the direction taken with NOVA, the Context Builder and the ComfyUI workflows developed by Axiris.

NOVA could structure the intent, the references, the composition and the visual rules of the project. The Context Builder would serve as the art direction layer. FLUX 3 could then execute that intent across several media.

The role of the language model would then no longer be merely to write a prompt, but to drive a genuine multimodal production chain.

A question of hardware

It is still too early to know whether the full backbone will run comfortably on a consumer graphics card.

An RTX 5090 32 GB might possibly suit a specialised, partially loaded or quantised version.

For the full model, an RTX 6000 Pro 96 GB seems more realistic today.

This remains a technical estimate by Axiris: BFL has not yet published any official configuration.

Axiris and the Early Access

Axiris's positioning matches particularly well the kind of use case being sought:

  • multimodal workflows
  • local integration
  • ComfyUI
  • structured art direction
  • reference management
  • visual continuity
  • interaction between LLMs and generative models

With NOVA and the Context Builder, Axiris does not present itself simply as a user wishing to test a new model, but as a studio already developing a genuine production architecture around these uses.

Key points

FLUX 3 Video enters Early Access. FLUX 3 Image will arrive next. FLUX 3 Dev is officially confirmed as open weights.

For Axiris, FLUX 3 above all confirms a shift already under way: the future will no longer be a succession of isolated models, but an environment capable of linking language, image, video, audio, references and art direction in a single creative system.