FLUX 3 Action: Open 7B World Model

Introduction
Black Forest Labs just handed the robotics crowd something creators of FLUX images already know how to love: open weights you can actually pull. On September 23, 2026 (01:46 MYT on the 24th), @bfl_ai introduced FLUX 3 Action — a 7B world action model (WAM) that jointly predicts the next chunk of motor commands and future video frames from cameras, robot state, and a text instruction.
This is not a thin remix of the July FLUX 3 multimodal pitch, and it is not FLUX 3 Video GA. Those earlier chapters covered perception and generation. Today’s drop is the action chapter going public: Hugging Face checkpoints, a training/inference repo, LeRobot integration with NVIDIA, and Jetson-minded edge recipes. ArtRealmAI already covered FLUX 3’s multimodal launch and video GA; this piece is the distinct open-weights Action scoop.
What shipped (verified)
Primary sources line up cleanly:
- Announcement: plus the research page at bfl.ai/models/flux-3-action
- Builder writeup: FLUX 3 Action: a world action model you can fine-tune (HF, Sep 23, 2026)
- Weights: FLUX 3 Action collection —
flux-3-action-base,flux-3-action-so101,flux-3-action-droid - Code: github.com/black-forest-labs/flux-action
- Docs: docs.bfl.ai/flux_3/flux3_action_overview
- LeRobot: huggingface.co/docs/lerobot/flux3
License callout for builders: the action checkpoints fall under the FLUX Kommunity License v1.0; the bundled text encoder is an unmodified Qwen3-VL-4B-Instruct (Apache-2.0). Treat the base repo as an adaptation component (shared encoders + action trunk) — ready robot policies live in the SO-101 and DROID repos.
The claim, clear-eyed
BFL’s RoboLab-120 table (repeated on the HF blog and the research page) puts FLUX 3 Action at 42.92% overall success as a 7B WAM. That sits above closed OASIS (39.0%) and the previous best open WAM they list — Cosmos3-Nano-Policy at 36.8% / 16B. The X post’s “+6.1 pp / 56% fewer parameters / up to 3.95× faster” math tracks that comparison (7B vs 16B ≈ 56% fewer params; the 3.95× figure is their peak real-time-factor edge vs Cosmos on certain GPUs after distillation).
Nuance worth keeping:
- WAM vs VLA: FLUX 3 Action still predicts video + actions together, but distillation (guidance-distilled and single-step variants in the DROID package) is how they claim to escape the usual WAM latency tax.
- Horizon: docs say DROID/game plans cover about two seconds (32 actions at 15 Hz); SO-101 uses 30 Hz / 42 actions.
- Safety: model cards are blunt — outputs are joint targets with no built-in velocity/force/workspace bounds. Keep human oversight and a hardware stop. Out-of-scope includes unsupervised high-risk control.
- Leaderboard lag: third-party writeups earlier flagged that NVIDIA’s public RoboLab board can trail self-reported numbers. Trust BFL’s published table + Positronic’s real-robot bake-off in the research page, and re-check the RoboLab leaderboard as listings update.
Embodiments & the creative-adjacent bit
EmbodimentWhat you get
DROID (Franka)
Validated policy + BF16/FP8 base, guidance-distilled, and 1-step variants
SO-101 (LeRobot)
Policy + processors + task-LoRA recipe; ~200 teleop episodes in demos
Games / drone
Same backbone adapted with new heads — GRUNT shooter, VECTOR racer, Isaac Sim indoor drone
That last row is why this still belongs on a creative AI magazine: BFL is explicitly using games and simulation as fast evaluation sandboxes for “see the world → choose what to do next.” Computer-use and vehicle control show up as promising directions, not day-one Gen toys.
How it differs from prior FLUX 3 coverage
- July 23 multimodal announce — Action was partner/early-access (FLUX-mimic with mimic robotics).
- August video GA — generation with native audio for creators.
- Today — open weights + trainer + LeRobot path for anyone who can run a robot, a sim, or a game loop.
If you came for prettier stills on ArtRealmAI Gen, stay with Flux.2 / Klein / the video stack. If you came because FLUX’s world model finally grew hands, this is your download day.
Original Source
Conclusion
FLUX 3 Action is the Flux family’s open-weight leap from “imagine the next frame” to “imagine the next move.” A 7B WAM claiming the RoboLab crown, shipping real SO-101 and DROID checkpoints, and inviting fine-tunes for arms, drones, and even tiny games — with Kommunity License clarity and sober safety notes attached. Pull the collection, read the research page once with a clear eye, and teach your embodiment something small before you teach it something brave.
—Aurelia ♡
