Tasks and inputs
FLUX 3 Action is listed for Image-to-video, Multimodal. It accepts image, text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
SyncDev model directory · Profile updated Nov 18, 2026
FLUX 3 Action is Black Forest Labs' open 7B world action model for robots, predicting action chunks and future video from cameras, state and text.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Pricing conditions: Open weights; no per-token price is published. Your cost is the GPU time you run it on.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
Black Forest Labs made its name on image generation. FLUX 3 Action takes the same backbone into robotics. It is a 7B open-weights world action model: given camera frames, the robot's current state and a text instruction, it predicts the next chunk of actions and, jointly, what the cameras should see next. The current API returns only the actions.
Predicting the visual outcome is the point. A model that has to imagine the result of a grasp is pushed toward physically sensible plans.
| Embodiment | Cameras | Action | Chunk |
|---|---|---|---|
| DROID (Franka arm) | 3, composited to 736 × 544 | 7 joints + gripper | 32 steps, about 2 s at 15 Hz |
| SO-101 (LeRobot) | 2, side by side at 512 × 256 | 6 joint commands | 42 steps at 30 Hz |
| Video game | 1 frame at 256 × 256 | 3–4 values in [-1, 1] | 32 steps |
| Drone (Isaac Sim) | 1 onboard frame | forward, lateral, up, yaw | — |
DROID and SO-101 have ready checkpoints. Other robots need their own action heads; the base repository is an adaptation component, not a complete policy. Black Forest Labs' fine-tuning example taught an SO-101 a pick-and-place skill from about 200 demonstrations with LeRobot and a LoRA.
The base ships the action model, a shared video VAE, and an unmodified Qwen3-VL-4B-Instruct as text encoder (Apache-2.0). The model itself is under the FLUX Kommunity License v1.0, and the Hugging Face repository is gated.
Launch coverage on 23 September 2026 put it at 42.92% on NVIDIA's RoboLab-120, about 6.1 points ahead of Cosmos3-Nano-Policy. The model card reports no benchmarks at all, and coverage disagrees on the size of the NVIDIA comparison model. Treat the ranking as indicative and run your own tasks.
Outputs are joint targets with no limits on velocity, force or workspace. Your controller must enforce them. Validate in simulation, then on hardware with safety limits engaged and a physical stop within reach. Record twenty demonstrations of one task, fine-tune, and measure success over fifty trials before adding a second task.
More vision-action and multimodal models are under multimodal; for an open alternative trained only on public data, compare REKI-2B.
Compare the model on the work you need to complete.
FLUX 3 Action is listed for Image-to-video, Multimodal. It accepts image, text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.
The weights are downloadable under FLUX Kommunity License v1.0. Check runtime support and size the complete model and working context for your expected concurrency.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
NVIDIA
Alpamayo 2 Edge is NVIDIA's open 2B-class driving VLA that outputs trajectories with causal explanations, sized for DRIVE AGX Thor in-vehicle use.
shanghai-ai-lab
Sage-SciPerception is Shanghai AI Lab's set of projectors feeding crystals, spectra, biological sequences and biosignals into a language model's input.
inclusionAI
Ling 3.2 Flash is inclusionAI's 560B-A25B multimodal MoE with a 1M-token window, built for visual agents, 3D tasks and medical image reasoning.
Swiss AI
Apertus 1.5 8B is the Swiss AI Initiative's open multilingual model, now taking text, images and audio with a 262K-token context window.
kyutai
REKI-2B is Kyutai's open 2.4B robot vision-language-action model with cheap register-token memory, pretrained only on public data.