BFL open-sources 7B world action model FLUX 3 Action, tops RoboLab 6.1pp ahead
bfl_ai · x · 2026-09-24
- Black Forest Labs released FLUX 3 Action, an open-weights 7B World Action Model that tops the RoboLab-120 leaderboard, beating the previous best open model by 6.1pp with 56% fewer parameters and up to 3.95x faster inference.
- It removes the WAM-vs-VLA trade-off: joint video-action prediction with a 2.13s action horizon (vs 1s for Pi0.5), processing robot motion 1.45–1.66x faster than Pi0.5; a guidance-distilled checkpoint runs 2.85–3.15x faster than the leading open WAM at a new SOTA success rate.
- Built on FLUX 3's image/video/audio pretraining with Self-Flow-derived compact representations and midtraining that jointly predicts actions and future frames; outputs 32 actions per step, predicts scene changes, and recovers from mistakes. Early results extend to games, simulated drones, and computer use.
- Weights, code, fine-tuning recipes, and benchmarks are open; natively integrated into Hugging Face LeRobot with NVIDIA, including Jetson edge deployment.
Related event: BFL Releases Open-Source 7B World Action Model FLUX 3 Action(3 posts)→
More from Embodied
- Black Forest Labs releases FLUX 3 Action: open 7B world-action model tops RoboLab-120 at 3.95x speed — robrombach · 2026-09-24
- How ACT works: a robot-learning primer on predicting 100 actions at once — DominiqueCAPaul · 2026-09-24
- π0.5 finetuning on real manufacturing: 1 hour of clean data beat the previous 17 — DominiqueCAPaul · 2026-09-24
- Tesla Optimus V3 leaked: everything we know so far — JOBhakdi · 2026-09-24
- Medical AI model xvr lands in Nature, aligning X-ray with CT/MRI in seconds — DrDatta_AIIMS · 2026-09-24
- Open-source robot arm controlled by Claude Opus 5 learns to paint the Golden Gate Bridge — KungRaLeo · 2026-09-24