Black Forest Labs releases open 7B FLUX 3 Action model

• Black Forest Labs released FLUX 3 Action as an open-weight 7B world action model, or WAM, in its launch post.
• The control loop jointly predicts scene changes and robot commands, then returns the next 32 actions, according to BFL's control-loop description.
• RoboLab-120 results span a 38.3% single-step checkpoint and a 42.92% leaderboard entry, while BFL's benchmark post reports faster robot-motion throughput than Pi0.5.
• The release includes backbone and embodiment-specific weights plus training recipes, as a team member's post confirms.

Video accounts for more than 95% of pretraining tokens, according to the technical report. The Apache-licensed GitHub package keeps training and BF16 or FP8 inference in separate transformer files.

F3A's public numbers are checkpoint-specific. The headline 42.92% leaderboard entry is 6.12 points above Cosmos 3 Nano's 36.8%, but it is not the single-step result.

• The single-step 7B checkpoint reached 38.3% ± 0.38, compared with 36.8% for Cosmos 3 Nano, according to the technical report.
• BFL's guidance-distilled checkpoint reached 42.2% ± 0.36 in the same report.
• The Hugging Face release guide lists FLUX 3 Action at 42.92% and Cosmos 3 Nano at 36.8% in its RoboLab table.

The report says BFL added randomly initialized action heads, fine-tuned on DROID, then evaluated on RoboLab. The scores therefore describe an embodiment-adapted policy, not zero-shot use of the pretrained backbone.

F3A conditions on recent camera frames, current state, and a task description. It produces a chunk of 32 actions alongside predicted future frames, then the system observes again and replans.

The loop has four pieces:

1. Visual observations from the robot's cameras.
2. State and language-task conditioning.
3. A joint prediction of actions and visual futures.
4. Re-observation after executing part of the action chunk.

BFL says this re-observe cycle can recover from mistakes during an attempt.

The model retains joint video-and-action prediction, the expensive property that distinguishes a WAM from a conventional action policy. BFL attributes its smaller backbone to Self-Flow representations and says its distillation removes a separate guidance pass while reducing sampling to one step.

The speed metric needs a small footnote: F3A's reported 1.34x to 2.28x advantage over Pi0.5 is robot motion processed per second, not per-call latency. The technical report attributes that distinction to F3A predicting a 2.13-second motion horizon versus Pi0.5's one second.

The release is a runnable policy stack, not just a checkpoint. The source package includes data preparation, distributed full fine-tuning, checkpoint resume, export, and inference.

• BFL released a DROID fine-tune and an SO-101 arm fine-tune.
• Both checkpoints are integrated with LeRobot, according to the Hugging Face guide.
• The SO-101 workflow uses a task LoRA through that LeRobot integration.
• BFL's launch post also describes NVIDIA Jetson edge deployment and fine-tuning recipes.

The release links point to the technical report and the weight download.

BFL also tested a hybrid controller where GPT 6 Astra delegates fast control to F3A and intervenes for harder tasks. In the report's simulation, that configuration solved 90% of episodes at $8.77 and eight minutes per success, versus $13.47 and 16 minutes for pure maximum-effort reasoning.

The [technical report linked from src:0|release material] frames the experiment as an early look at combining a frontier reasoner with a low-latency policy.

BFL trained task-specific policies for games and a simulated drone, calling the results early but promising. Its report uses gaming as a navigation testbed for visually driven computer-use agents.