Black Forest Labs (BFL) has introduced FLUX 3, a multimodal foundation model capable of generating audio-visual content and driving physical robots, according to a blog post on July 23, 2026. The company collaborated with mimic, a robotics firm, to develop FLUX-mimic, a video-action model tested and deployed at Audi production lines. FLUX 3 is trained jointly across image, video, and audio data, with video prediction accounting for over 95% of total compute costs. Audio prediction makes up less than 0.5% of the data tokens in 720p video, while action prediction is treated as another modality related to the physical reality represented in video.

During large-scale training, integrating action prediction initially reduced human ratings on text-to-video and image-to-video by up to 10%. After 3500 training steps, the model regained its full quality while simultaneously gaining the ability to predict actions. FLUX-mimic uses a lightweight action decoder trained on intermediate features from the FLUX 3 backbone. This architecture relies on Self-Flow, a framework that unifies generation and representation learning to produce high-quality, disentangled world models.

Benchmarks indicate that FLUX-mimic's action decoder outperforms previous vision-language-action models, achieving state-of-the-art success rates when the backbone is fine-tuned. The model exhibits high sample efficiency, with some experiments requiring half the training steps compared to models without Self-Flow. In real-world factory tasks, the system handles kitting, inserting electronic control units, and manipulating flexible materials like seals and cables. The FLUX-mimic backbone is optimized to produce world representations in under 80ms on an NVIDIA RTX 5090 GPU. Combined with an optimized deployment stack, the system achieves a total reaction time of 101ms, which mimic and BFL claim is on the same order of magnitude as human visual reaction time. Audi is currently testing these robots for soft-body manipulation tasks that were previously difficult for conventional automation to address.

Black Forest Labs (BFL) has introduced FLUX 3, a multimodal foundation model capable of generating audio-visual content and driving physical robots, according to a blog post on July 23, 2026. The company collaborated with mimic, a robotics firm, to develop FLUX-mimic, a video-action model tested and deployed at Audi production lines. FLUX 3 is trained jointly across image, video, and audio data, with video prediction accounting for over 95% of total compute costs. Audio prediction makes up less than 0.5% of the data tokens in 720p video, while action prediction is treated as another modality related to the physical reality represented in video.

During large-scale training, integrating action prediction initially reduced human ratings on text-to-video and image-to-video by up to 10%. After 3500 training steps, the model regained its full quality while simultaneously gaining the ability to predict actions. FLUX-mimic uses a lightweight action decoder trained on intermediate features from the FLUX 3 backbone. This architecture relies on Self-Flow, a framework that unifies generation and representation learning to produce high-quality, disentangled world models.

Benchmarks indicate that FLUX-mimic's action decoder outperforms previous vision-language-action models, achieving state-of-the-art success rates when the backbone is fine-tuned. The model exhibits high sample efficiency, with some experiments requiring half the training steps compared to models without Self-Flow. In real-world factory tasks, the system handles kitting, inserting electronic control units, and manipulating flexible materials like seals and cables. The FLUX-mimic backbone is optimized to produce world representations in under 80ms on an NVIDIA RTX 5090 GPU. Combined with an optimized deployment stack, the system achieves a total reaction time of 101ms, which mimic and BFL claim is on the same order of magnitude as human visual reaction time. Audi is currently testing these robots for soft-body manipulation tasks that were previously difficult for conventional automation to address.