Black Forest Labs x mimic robotics: FLUX-mimic runs video generation and robot actions on one backbone
On July 23, Black Forest Labs (BFL) released FLUX 3, and mimic robotics released FLUX-mimic the same day. Both are two landing points of the same architectural judgment.
FLUX 3: video prediction as world modeling
FLUX 3 is a multimodal foundation model with image, video, and audio jointly trained on the same backbone from the start. BFL published a striking compute-allocation figure: video prediction accounts for over 95% of total training compute. The reasoning is straightforward: to generate believable video, the model must learn contact, motion, weight, and causality — any mistake makes the video look wrong. Audio takes less than 0.5% of 720p video tokens, because once video prediction is learned, the video-audio causal relations (footsteps, door closures, lip sync) come almost for free.
This is another implementation path for Yann LeCun's argument that LLM-only approaches are insufficient and models must learn world dynamics — except BFL approaches it from the video side rather than the language side.
FLUX-mimic: a Video-Action Model (VAM)
BFL gave mimic robotics early access to FLUX 3, and mimic fine-tuned it for robot control, calling the architecture a Video-Action Model (VAM). The core claims:
1. Actions are a low-dimensional representation of robot state, tightly coupled to visual observation. Video, audio, and action frames are partial observations of the same physical reality. Once the model learns the physics behind video and audio, action prediction is not a new task — it is another view of a world it already models.
2. Action training does not permanently harm video generation quality. BFL added action prediction as part of a curriculum during large-scale training. Human preference scores for text-to-video and image-to-video initially dropped about 10%, but after 3,500 steps the model recovered full video generation quality while also learning to predict actions. Adding the action modality is not a new direction; it is an extension of the same backbone.
3. FLUX-mimic is not a research demo — Audi deployed it on real production lines. Tasks include placing parts into trays, inserting ECUs into fixtures, assembling components, and handling flexible seals and cables — soft-body operations that traditional automation cannot do. This builds on mimic's existing hardware stack: the mimic hand M1, the wearable U1, and their frontier middleware.
The numbers
- Backbone latency on a single RTX 5090: under 80 ms
- End-to-end robot reaction time: 101 ms
- Data efficiency: mimic claims VAM improves on traditional VLA by 10x — because, in their framing, "video pre-training is far cheaper than robot data."
My take
BFL this time is not just "another video model" — it takes the thesis "video prediction = physical world modeling" all the way to real-robot deployment. In 2026 H2, embodied competition may shift from "dedicated action model + separate world model" to "one backbone outputting both pixels and joint angles." This differs from VLA routes like LingBot-VA, π0, and Helix: VLAs treat vision-language as pre-training, while VAM treats video generation as pre-training. The latter bets that video prediction is harder than image-language alignment, and that mastering it means mastering world dynamics.
Limitations
1. BFL's reported preference rates range from 52% to 93%, but sample sizes and evaluation pools are unspecified — the data is preliminary. 2. The real-robot deployment partner is Audi Production Lab, on relatively structured industrial tasks (tray loading, ECU insertion) — not general-purpose home robotics. 3. No API, parameter count, pricing, or GA timeline given; open weights for FLUX 3 Dev are promised "later this year." 4. mimic-video is mimic's own prior work; VAM did not appear from nowhere — this is a retraining with FLUX 3 as the backbone.
Overall: a new playbook for physical AI in 2026 H2 — treating the training compute of a generative video model (95% of FLUX 3's compute) as physical world modeling compute. The same training budget yields two commercial outlets: creative tools and industrial robots. It is LeCun's world-model thesis in its capital-markets version, not the paper version.