FLUX 3 Action: A 7B World Action Model for Robot Control

TL;DR

FLUX 3 Action (F3A) is BFL’s open-weight 7B World Action Model derived from the multimodal FLUX 3 backbone (FLUX 3 — Real World Models: Towards Multimodal Flow Models as the Backbone of Visual Intelligence), positioned as a Pareto move that keeps joint video+action denoising (the WAM property that transfers world understanding to control) while cutting the compute cost that made WAMs slow. On the RoboLab-120 leaderboard, a single-step 7B checkpoint reports 38.3% ± 0.38 success — beating Cosmos 3 Nano’s 36.8% at less than half the parameters — while a guidance-distilled variant pushes the state-of-the-art to 42.2% ± 0.36. The distillation stack combines multimodal Self-Flow pretraining + video-action midtraining + DROID joint finetuning + guidance distillation + single-step distillation + FP8, delivering 1.34–2.28× speedup over π0.5 on workstation/datacenter GPUs and 1.52–3.95× over Cosmos 3 Nano at similar or better task success. BFL also reports a hybrid-with-reasoner result: delegating from GPT-6 Astra to F3A only when needed reaches 90% success at $8.77 / 8 minutes per success, 29% cheaper and 40% faster than pure Astra reasoning ($ 13.47 / 14 min) — where the same delegation experiment with π0.5 as the fast policy barely moves the needle.

Key claims

FLUX 3 Action reaches 42.92% success on the RoboLab leaderboard vs 36.8% for Cosmos 3 Nano (16B), 33.3% for BiMind v0.1, 28.0% for π0.5, and 25.7% for DreamZero (14B) — the reported top of the open-policy board at 7B [Table (RoboLab leaderboard)].
Video prediction accounts for >95% of pretraining tokens; the resulting checkpoint reaches 11.6% DROID→RoboLab success from randomly-initialized action heads, vs <1% without pretraining — video-heavy multimodal pretraining is load-bearing for downstream action prediction [§Pretraining and Midtraining, Fig. 4].
Midtraining on a mix of joint video/audio (37.0% samples), gaming (19.5%), egocentric human hands (13.5%), handheld grippers (14.0%), and teleop across 14 embodiments (15.9%) with an EE50 relative-pose action head reaches 17.3% RoboLab success without any additional DROID finetuning [§Pretraining and Midtraining, Table 2, Fig. 4].
Joint video+action denoising has separate signal-to-noise transition regions per modality (σ(t) ≈ 3.3, t ≈ 0.964); training with a logit-logistic timestep distribution shifted to α = 42 with action scaling s = 2 consistently outperforms alternatives across seeds on RoboLab simple tasks [§Noise-to-Signal Transition, Fig. 5–7].
Classifier-free guidance can be split per modality on a WAM; a grid search picks Video CFG = 4.0 and Action CFG = 1.0 as best (42.00% success on the 120-task suite) — video benefits from guidance, actions do not [§Split Guidance, Fig. 11].
Guidance distillation removes the second forward pass without loss (0.6–1.08 pp improvement) and yields 1.8–2× speedup; step distillation from 4 to 1 sampling step yields an additional 3.15–4× speedup at a 3.51–4.32 pp cost — the 1-step distilled model still beats Cosmos 3 Nano by 1.12–1.88 pp while running 9.19–13.22× faster [§Inference Efficiency].
Online Power EMA with σrel = 0.10 consistently beats traditional EMA (β = 0.999) and outperforms post-hoc reconstruction from stored Power-EMA snapshots — online-during-training averaging is the recommended DROID recipe [§EMA Analysis, Fig. 12].
On the Su et al. (2026) hard-task benchmark, no pure action policy solves every task, but GPT-6 Astra with maximum reasoning does — at 35.77 s reasoning per second of robot motion, $13.47 and 16 min per success; F3A alone runs at 21.08 ms per second of robot motion but doesn’t solve every task [§Action Policies].
A hybrid where GPT-6 Astra delegates control to F3A and only reasons when needed reaches 90% task success at $8.77 and 8 min per success — 29% cheaper and 40% faster than pure Astra reasoning; the same delegation strategy with π0.5 as the fast policy barely improves ($ 12.28 vs $13.47) [§Action Policies].
In FP8, F3A step-distilled model runs at 32.29 ms per prediction (32 actions covering 2.13 s of motion at 15 Hz), vs 320.40 ms for Cosmos 3 Nano FP8 and ~32.00 ms for π0.5 BF16 (10 steps, guidance off) [Tables of workload latencies].

Method

FLUX 3 Action is a World Action Model — joint video+action prediction in a single 7B multimodal backbone — obtained by finetuning FLUX 3 for action prediction. Training goes through four phases.

Pretraining. The FLUX 3 backbone is trained on a mix of image, video, and audio data with video >95% of training tokens under the Self-Flow (Self-Flow: Self-Supervised Flow Matching for Scalable Multi-Modal Synthesis) objective. Load-bearing per the paper: pretraining alone lifts DROID→RoboLab from <1% to 11.6% success with randomly-initialized action heads.

Midtraining. An action-focused phase mixes 37% joint video/audio, 19.5% gaming (64-dim mouse+keyboard action space), 13.5% egocentric human hands, 14% handheld grippers, and 15.9% teleop across 14 embodiments — with the EE50 relative-pose action head (25-dim per end-effector: 3 translation, 6 rotation-as-first-two-rows of R matrix following Zhou et al. 2019, 16 hand/gripper). EE50 evaluation on RoboLab without any DROID finetuning reaches 17.3%.

DROID finetuning. Adds joint video+action heads on the midtrain checkpoint. For randomly-initialized action heads (Joint recipe), midtraining weights are frozen for 1k steps then warmed up linearly to 2e-4 over 2k steps; action-head LR warms up to 1e-3. EE50 finetuning (starting from midtrained action heads) uses a single warmup. Joint finetuning at 4× batch size (2048) overtakes EE50 finetuning after 10k steps and reaches 40.13% vs 37.88% at the end of training. The best checkpoint is picked from a 5k-step linear-LR-cooldown tail (40% at 26k steps).

Distillation for inference. Two stages both operate on the joint video+action model. Guidance distillation freezes the DROID checkpoint as teacher and trains a student to produce the guided prediction directly, removing the second CFG pass at inference (1.8–2× speedup, 0.6–1.08 pp gain). Step distillation initializes from the guidance-distilled checkpoint and trains a single-step student under the teacher’s guidance, cutting sampling steps 4→1 (3.15–4× speedup, 3.51–4.32 pp drop). Both stages preserve joint video+action prediction — the imagination path stays.

Inference optimization. Row-wise dynamic FP8 (+0.05 pp to -0.76 pp vs BF16), prompt-token KV caching across timesteps, static-shape torch.compile with reduced-overhead mode, and moving step-invariant computation out of blocks into shared prep.

Training-time detail worth flagging: separate S-shaped signal-to-noise transition envelopes for video and action modalities motivate a logit-logistic timestep distribution shifted with α = 42 (mode near action transition) and an action-scale s = 2 that pushes the joint model toward approximating p(m₁ | s·m₂) p(s·m₂) — video is denoised first, actions conditioned on it.

Results

Task success (RoboLab, full 120-task suite):

FLUX 3 Action guidance-distilled: 42.92% (SOTA on the leaderboard)
OASIS WAM (closed): 39.0%
Cosmos 3 Nano-Policy (16B, open): 36.8%
Phoenix (closed): 34.4%
BiMind v0.1 (closed): 33.3%
π0.5 (3.3B, open): 28.0%
DreamZero (14B, open): 25.7%
Cosmos 3 Edge-Policy (4B): 22.9%
π0-FAST (3B): 15.5%
GR00T N1.6 (3B): 7.2%
π0 (3.3B): 5.0%

F3A internal ablations (RoboLab, mean ± SEM):

Base BF16 (4 steps, guidance on): 41.60 ± 0.46% (6 seeds)
Base FP8: 41.16 ± 0.30% (8 seeds)
(Guidance-distilled and step-distilled numbers reported with matched precision in the tables — table truncated at 30KB fetch; the “up to 42.2% ± 0.36” claim in the executive summary corresponds to the guidance-distilled variant.)

Split-guidance grid (120 tasks): best cell at Video CFG=4.0 / Action CFG=1.0 → 42.00%; the gradient prefers higher video guidance and lower action guidance.

Latency (per prediction, 32 actions = 2.13 s of motion at 15 Hz):

F3A 4 steps guidance on: 246.42 ms BF16 / 182.00 ms FP8
F3A 4 steps guidance off (guidance-distilled): 136.25 / 101.71 ms
F3A 1 step guidance off (step-distilled): 41.06 / 32.29 ms
Cosmos 3 Nano 4 steps guidance on: 387.71 / 320.40 ms
π0.5 10 steps guidance off: 31.99 ms BF16 (FP8 not evaluated)

Real-time factor (RTF = processing / robot motion): in FP8, F3A step-distilled provides 1.34–2.28× speedup over π0.5 BF16 on workstation and datacenter GPUs; on RTX 5090 consumer hardware, F3A is 1.15–2.42× slower than π0.5 in FP8/BF16 respectively. FP8 F3A base + guidance-distilled runs 1.52–3.95× faster than FP8 Cosmos 3 Nano across consumer/workstation/datacenter GPUs.

Hybrid reasoning + policy (Su et al. 2026 benchmark):

GPT-6 Astra alone (max reasoning): solves all tasks, 35.77 s reasoning per s of motion, $13.47 and 16 min per success
F3A alone: 21.08 ms per s of motion, ~$0.09 and <2 min per success on H200, doesn’t solve every task; could serve 47 concurrent rollouts at no additional cost
Astra + π0.5 hybrid: $12.28 and 13.5 min per success (barely improves)
Astra + F3A hybrid: 90% success at $8.77 and 8 min per success — 29% cheaper, 40% faster than pure reasoning

Pretraining/midtraining ablation (DROID→RoboLab success, small-batch protocol):

No pretraining: <1%
Pretraining only, randomly-initialized action head: 11.6% → 12.4%
Midtraining, EE50 head, no DROID finetune: 7.2% → 14.7% → 17.3%
Midtraining, Joint DROID finetune: 2.3 pp lift over last pretraining checkpoint, rising to 18.6% at late midtraining

Why it’s interesting

F3A is the second BFL production landing on the World Foundation Models “WFM-as-policy-internal-substrate” pattern after FLUX 3 x mimic: The Next Generation of Video-Action Models FLUX-mimic, but it takes the opposite architectural choice: where FLUX-mimic freezes the FLUX 3 backbone and reads action decisions from intermediate features via a small decoder (VAM factorization, mimic-video recipe), F3A keeps joint video+action denoising end-to-end and pays the compute back with a stacked distillation pipeline. This directly answers the WAM-vs-VLA tradeoff Pretrained to Imagine, Fine-Tuned to Act: The Rise of World-Action Models frames — you don’t have to give up joint video prediction to hit VLA-class inference latencies. On the VLA Models recipe-lever board, F3A slots next to Fast-WAM: Do World Action Models Need Test-time Future Imagination? Fast-WAM’s “video co-training during training, no future-video denoising at inference” middle position, but with the imagination path preserved and distilled rather than removed — a distinct point on the same axis.

The hybrid Astra + F3A result is the sharpest concrete cash-out yet of the frontier-LLM-as-controller + fast-policy-as-executor thread running through the wiki (Robot-Use Agents robot-use agents, Jay Chooi (Robocurve) — Astra vs Fable 5.1 on real-robot manipulation via IK Astra-vs-Fable via IK, Agent as Policy for Robotic Manipulation AGP): the fast policy’s quality determines whether hybrid delegation actually saves money — π0.5 gets you $12.28/success vs pure reasoning's$ 13.47 (barely), F3A gets you $8.77 (materially). It also validates the Diffusion Distillation stack (guidance distillation + step distillation + FP8) as the specific set of moves that convert a large multimodal backbone into a control-loop-viable policy, complementing Sol-Attn: Accelerating Video Generation Inference via On-the-Fly Attention Sparsification‘s on-the-fly sparsification and C³ache: Accelerating World Action Models with Cross Inference Chunk Cache‘s cross-chunk caching on the same latency axis.