Black Forest Labs’ FLUX 3 Action report positions a 7B open-weight world-action model — camera, state, and text in; joint video+action prediction out — as a Pareto move on RoboLab versus larger open WAMs and smaller VLAs.
What BFL and Hugging Face published
Per BFL, FLUX 3 Action is derived from the multimodal FLUX 3 backbone and pretrained with a strong emphasis on video. The Hugging Face post states it takes a camera frame and a text instruction and returns the next 2 seconds of actions; BFL’s control table lists 32 actions at 15 Hz for a 2.13s full-chunk duration (versus Pi0.5’s 15 actions / 1.0s). Video and action tokens share a single noise level and are denoised jointly.
Embodiment recipes cover DROID finetuning and an SO-101 path. Both checkpoints integrate into LeRobot with parameter-efficient fine-tuning recipes. Weights sit in the FLUX 3 Action collection under the FLUX Kommunity License v1.0 — cite that license name carefully and read the full text before commercial use.
“FLUX 3 Action is an open weights 7B world action model. It takes a camera frame and a text instruction and returns the next 2 seconds of actions.”Hugging Face — FLUX 3 Action blog, September 23, 2026
RoboLab and efficiency (from the primaries)
On RoboLab-120, BFL reports FLUX 3 Action at 42.92% overall success as a WAM — ahead of Cosmos3-Nano-Policy at 36.8% (16B) and π0.5 at 28.0% (3.3B VLA). Guidance-distilled and step-distilled FP8 variants are detailed for latency trade-offs. Attribute every SR% and speedup figure to BFL’s tables only — this desk did not re-run RoboLab.
| Line | Value | Compare | Notes |
|---|---|---|---|
| Parameters | 7B | Cosmos 3 Nano 16B | Open WAM |
| RoboLab SR% | 42.92% | 36.8% Cosmos · 28.0% π0.5 | BFL table |
| Actions / chunk | 32 @ 15 Hz | Pi0.5: 15 @ 15 Hz | ~2.13s vs 1.0s |
| Embodiments | DROID + SO-101 | — | LeRobot integration |
| License | FLUX Kommunity License v1.0 | Read full text | |
What Alberta operators should do
For industrial robotics Alberta shops and UAlberta / applied-robotics paths, the operator move is to treat FLUX 3 Action as an open WAM candidate you can fine-tune on a small teleop set via LeRobot — not as a drop-in AGI robot. Start with SO-101 or DROID-style pick-and-place demos, measure success on your own fixture, and keep the Kommunity license review in the same packet as the weights download.
Pair this briefing with Opcelerate’s AI consulting Alberta path and private AI security lane, and keep reading the city desk at The Super Intelligence Times. Open weights on a workshop GPU still need a data-governance story when plant video or proprietary CAD leaves the floor.
The guardrail
RoboLab 42.92% is BFL’s reported figure; hybrid Astra-delegation cost tables on the BFL page are lab economics, not Alberta MSSP quotes. Do not invent retail GPU or per-success prices for local decks. Read the FLUX Kommunity License v1.0 before any commercial deployment, and do not treat a 2-second action horizon as full autonomy.