FLUX 3 Action: 7B Open-Weight World Action Model Beats 16B Cosmos 3 Nano on RoboLab-120

Forum topic · 小凯 · 2026-09-25

Summary

On September 23, 2026, Black Forest Labs released FLUX 3 Action, an open-weight 7B World Action Model (WAM) that jointly predicts robot actions and future video frames. It achieves a 42.92% task success rate on the NVIDIA RoboLab-120 simulation benchmark, surpassing NVIDIA's 16B Cosmos 3 Nano (36.8%) and Physical Intelligence's 3.3B π0.5 (28.0%), while running 1.52-3.95x faster. Code, weights, and training recipes are available on Hugging Face and GitHub, with fine-tuned variants for DROID and SO-101 embodiments and LeRobot framework integration. The technical report details three engineering choices: noise-schedule matching between video and action via a logit-logistic distribution, freezing the backbone for the first 1k fine-tuning steps, and asymmetric Split Guidance (video CFG 4.0, action CFG 1.0). In 30 real-robot trials on a Franka arm, it completed 28/30 (93.3%). A hybrid architecture pairing FLUX 3 Action with the GPT-6 Astra reasoning model reached 90% success at 29% lower cost and 40% faster than pure reasoning inference. Open questions remain around commercial license terms, cross-embodiment transfer costs, and the absence of built-in safety limits on raw joint outputs.

On September 23, 2026, Black Forest Labs (BFL) released FLUX 3 Action, an open-weight 7B World Action Model (WAM). It reached a 42.92% task success rate on the NVIDIA RoboLab-120 simulation benchmark, beating Cosmos 3 Nano's 36.8% (16B parameters) and π0.5's 28.0% (3.3B parameters). Code, weights, and training recipes were released simultaneously on Hugging Face and GitHub, with fine-tuned variants for the DROID and SO-101 embodiments and integration into the LeRobot framework.

Key takeaway

FLUX 3 Action proves the "video prediction = world modeling" thesis on real robots. With less than half the parameters of Cosmos 3 Nano, it achieves a higher task success rate while running 1.52-3.95x faster. Its main engineering significance: when the backbone has already learned world dynamics from video pre-training, robot policies do not need to learn actions from scratch.

WAM vs. VLA

Robot policy models currently split into two schools:

| School | Representative models | Core claim | Key capability | |---|---|---|---| | VLA (Vision-Language-Action) | π0 / π0.5 / GR00T N1.6 | Vision-language pre-training is the bridge to robot policies | Natural-language task descriptions | | WAM (World Action Model) | Cosmos 3 / FLUX 3 Action / DreamZero | Video prediction is the bridge to robot policies | Jointly predicts future frames and actions |

The WAM hypothesis is more aggressive: a model that has learned "how objects move" from massive video data has implicitly grasped physics, and only needs light robot-data fine-tuning to translate that knowledge into joint commands. FLUX 3 Action is currently the strongest open-source example of this route.

The numbers

| Model | Type | Params | RoboLab-120 SR% | Notes | |---|---|---|---|---| | FLUX 3 Action (baseline FP8) | WAM | 7B | 42.92% | Best open-source | | FLUX 3 Action (guidance-distilled FP8) | WAM | 7B | 42.24% | 1.8x-2x faster | | FLUX 3 Action (step-distilled) | WAM | 7B | 37.92% | Single-step, 3.15x-4x faster | | Cosmos 3 Nano | WAM | 16B | 36.8% | NVIDIA open-source | | OASIS | VLM+WAM | Closed | 39.0% | Closed-source | | Phoenix | TAMP+FM | Closed | 34.4% | Closed-source | | BiMind v0.1 | VLA | Closed | 33.3% | Closed-source | | π0.5 | VLA | 3.3B | 28.0% | Physical Intelligence, open-source | | DreamZero | WAM | 14B | 25.7% | Open-source | | GR00T N1.6 | VLA | 3B | 7.2% | NVIDIA open-source |

Key observation: 7B FLUX 3 Action outperforms 16B Cosmos 3 Nano by six points — a large gap within the WAM route, empirically supporting "smaller model + better video pre-training > bigger model + mediocre policy training" on RoboLab.

BFL also disclosed 30 real-robot trials (Frank arm, executed by third-party Positronic Robotics): FLUX 3 Action completed 28/30 (93.3%), Cosmos 3 Nano 27/30 (90.0%), DreamZero 20/30 (66.7%), π0.5 13/30 (43.3%).

The control loop

FLUX 3 Action predicts 32 actions at once (~2.13 s of motion, 15 Hz). Inputs:

One or more camera images (arranged per embodiment: DROID's three 544×736 cameras, SO-101's two side-by-side cameras, or 512×512 game frames)
State vector (joint positions or last executed actions)
Language task description (frozen Qwen3-VL-4B encoder)

Outputs:

32 actions (joint commands)
Future video frames (jointly denoised at the same noise level as actions)

The control loop is "observe → think → act → re-observe":

1. The model receives the current camera frame, state, and task description 2. It jointly predicts 32 actions and future video 3. The robot executes part of the actions 4. Re-observe and repeat

Re-observation is the WAM's key differentiator: it can catch deviations mid-execution and adjust subsequent actions based on new observations, whereas pure VLA models typically only re-observe after all 32 actions are executed, making error correction more costly.

Three engineering innovations in the training recipe

1. Noise-schedule matching between actions and video. Joint prediction requires denoising video and actions at the same noise level, but the two have different signal strengths (action signals typically saturate later than video). Using minimum-envelope loss analysis, BFL found the action signal transitions later than video, and adopted a logit-logistic distribution (scale=0.75, shift α=42) plus an action scaling factor s=2 to optimize joint denoising trajectories.

2. Freezing the backbone for the first 1k steps of DROID fine-tuning. After randomly initializing the action head, it is first warmed up for 1k steps to adapt to the backbone's feature space, then the learning rate is linearly ramped to 2e-4, followed by 2k steps of joint training — a stable "warm the head first, then the whole body" recipe.

3. Split Guidance. Video and actions use separate classifier-free guidance, with the optimal combination being video CFG=4.0 and action CFG=1.0. The asymmetry suggests video generation needs stronger guidance for realism, while action generation needs only light guidance to preserve responsiveness.

Distillation and deployment

BFL released two distillations:

Guidance distillation: collapses two forward passes into one, 1.8x-2x speedup, with performance actually improving 0.6-1.08 points
Step distillation: reduces sampling to a single step, 3.15x-4x speedup, with a 3.51-4.32 point performance drop

For deployment, BFL provides NVIDIA Jetson edge inference support. FLUX 3 Action runs a single action-block inference in 41.06 ms on a B200 GPU — fast enough to close a real control loop rather than being "too slow to think to be useful."

Hybrid with reasoning models

BFL tested a hybrid architecture combining FLUX 3 Action with the GPT-6 Astra reasoning model: Astra handles complex planning while FLUX 3 Action handles fast low-level control. Simulation results:

| Configuration | SR | Cost/success | Time/success | |---|---|---|---| | GPT-6 Astra (xhigh) | 100% | $13.47 | 16m23s | | FLUX 3 Action (pure) | Partial | $0.087 | ~1.75 min | | Hybrid (Astra low effort + F3A) | 90% | $8.77 | 8m08s |

The hybrid is 29% cheaper and 40% faster than pure reasoning while retaining 90% success — pointing to a clear engineering path: expensive-but-precise reasoning models for planning, fast-and-cheap policy models for execution.

Whom to trust

BFL's self-reported numbers (42.92% on RoboLab, 28/30 real-robot) come from the official technical report and Hugging Face model card. The third-party RoboLab benchmark is maintained by NVIDIA, and comparison figures for Cosmos 3 Nano etc. are NVIDIA self-reported. The commercial terms of the FLUX Kommunity License v1.0 have not been fully published — the detail the open-source community cares about most. BFL's comparisons cover only open-source models and exclude closed systems like Google RT-2 and Physical Intelligence Helix.

Unanswered questions

1. Commercial license boundaries. Both non-commercial and commercial options exist, but pricing, restrictions, and redistribution terms are unpublished — the biggest uncertainty for teams building products on FLUX 3 Action.

2. True capability of closed systems. Physical Intelligence's self-reported 28.0% for π0.5 is the open-source version; closed-source and newer versions are undisclosed. Benchmarks against Google RT-H and Figure Helix are also missing. FLUX 3 Action's 42.92% is "best open-source WAM," not "best policy model overall."

3. True cost of cross-embodiment transfer. BFL provides fine-tuned versions for DROID and SO-101, but transfer costs between embodiments (Franka to UR, 6-DOF to 7-DOF), performance loss, and whether relabeling is required remain unspecified.

4. Safety and boundary conditions. The model outputs raw joint targets with no built-in velocity, force, or travel limits. BFL's own technical report states this and leaves safety layers to users — an engineering hazard for consumer and home robotics.

---

Sources: BFL official technical report (FLUX 3 Action, Sept 23), Hugging Face model card (FLUX 3 Action collection), Black Forest Labs HuggingFace Blog (Sept 23), BFL GitHub repo (flux-action), Cosmos 3 paper arXiv:2606.02800, π0.5 paper arXiv:2504.16054, Self-Flow paper arXiv:2603.06507, RoboLab paper arXiv:2604.09860, plus coverage from DataNorth AI, The Decoder, AI Primer, and AI and Tech News.