Introducing FLUX 3 Action. An open weights 7B World Action Model that achieves first place on the RoboLab benchmark. It outperforms the previous best open model by 6.1 percentage points while using 56% fewer parameters and running up to 3.95x faster.⁠⁠ FLUX 3 Action removes the usual trade-off between... world action model performance and VLA speed: it still predicts video and actions together, but plans more than twice as far ahead and runs faster per second of robot motion than the strongest open VLA. Teams can fine-tune FLUX 3 Action on their own demonstrations to create policies for a particular robot and task. Together with NVIDIA, we also integrated FLUX 3 Action natively into Hugging Face's LeRobot, with fine-tuning recipes included and edge deployment on NVIDIA Jetson. Beyond robotics, we’re also seeing promising results training task-specific policies for acting in simulated environments like gaming, controlling a vehicle, computer use, and wherever else a model needs to understand a visual environment and then choose what to do next. FLUX 3 Action builds on the same image, video, and audio pretraining as FLUX 3, but uses a smaller architecture designed for practical deployment. In midtraining, we trained the model to predict actions and future frames together. We’re releasing the weights, code, fine-tuning recipe, benchmarks, and reproducible examples so researchers and developers can build on the model with their own robots, environments, and tasks (see below).show more

Read the full blog: Download the weights: Talk to our robotics team:

FLUX 3 Action removes the usual trade-off between the higher success rates of WAMs and the speed of VLAs. Its single-step 7B checkpoint outperforms every other open policy on RoboLab, while processing each second of robot motion 1.45× to 1.66× faster than Pi0.5 (the strongest open VLA model). It retains the core WAM architecture, predicting video and actions jointly, but covers a longer 2.13-second action horizon compared with 1 second for Pi0.5. When latency matters less, our guidance-distilled checkpoint raises the state-of-the-art success rate on RoboLab while running 2.85× to 3.15× faster than the previous leading open WAM.

FLUX 3 Action builds on the same image, video, and audio pretraining as FLUX 3, but uses a smaller architecture. More efficient representations learned through our research on Self-Flow made this smaller size possible. In midtraining, we trained the model to predict actions and future frames together.

FLUX 3 Action takes recent camera frames, the system’s current state, and a description of the task. It then returns the next 32 actions while also predicting how the scene will change. The system repeatedly observes, plans, acts, and adjusts.⁠⁠ The model is also capable of recovering from its own mistakes.

Beyond robotics, FLUX 3 Action showed promising early results when we trained task-specific policies for playing games and controlling a simulated drone — suggesting the same approach could apply wherever a model needs to understand a visual environment and choose what to do next (e.g. simulations, games, and computer use).

FLUX 3 Action controlling a simulated drone to take a specific requested action.

More details here:

Not the weights we want but ok. When are the Flux 3 Video weights dropping?

when open weights video! not action and robotics!

As cool as this is, what about releasing the Flux 3 video weights like you guys mentioned? I am pretty sure thats what everyone really wants to know about 😂😂

But what about Flux 3 video weights?

And what about Flux 3.0?

Flux videos pls open weight

Nice!

@victormustar Live BFL!

flux 3 open weights never came , you guys are liars

7B parameters outperforming larger models is seriously impressive engineering

We need FLux 4 for image and video

And what about... Flux 3 Image? Did you forget?

🔥

The 6.1pp at 56% fewer params is the leaderboard flex; the real question is whether twice-ahead video+action still holds once teams leave RoboLab and fine-tune on their own Jetson demos.

The model looks really promising. FLUX is already a great backbone. It definitely needs to be tested across more domains, though.

your world model never left the render

For a moment I thought 7B parameter video generation model 🙂

@Presidentlin

7B parameters is a featherweight. I’m curious how it handles complex spatial reasoning in cluttered scenes compared to those bloated 10B+ rivals. If the efficiency holds up at scale, that open-weight flexibility could be seriously handy for local setups

this is the way

Open-weight robotics AI is moving fast. 3.95× faster is a huge improvement.

gz on #1

The 6.1-point gain with 56% fewer parameters is a compelling result, but the speedup may be the bigger win for real robots: faster action loops can matter more than a small offline score bump. Fine-tuning on task-specific demonstrations also makes the open-weights release much more usable than a fixed benchmark model.

Fewer parameters and faster inference make this robotics result stand out.

🤯. When Flux 3 video and image Open Source finetunes? So desparately waiting for them to drop.

smaller faster stronger, pick all three apparently

@grok 它消费了什么样子的数据集 ？ 数据从谁那边买的

robots getting faster while I still trip over flat ground

The way this model predicts video and actions together while staying so fast is a massive leap for robotics. We actually went deeper on this here:

画图厂长手了

7B on a Jetson is the part robotics people will care about. What control rate does it actually hit on-device?

okay thats fucking sick