FLUX 3: The Multimodal Foundation Model That Learns from Video, Audio, and Images to Control Robots

Black Forest Labs (BFL) introduced FLUX 3 on July 23, 2026 — a multimodal foundation model that jointly learns from images, video, and audio within a unified architecture, and extends that understanding to predict robotic actions . It is BFL's first model built for "real-world visual intelligence": AI systems that perceive, predict, and act across both digital and physical environments .

Key Capabilities

Video generation — FLUX 3 can generate up to 20-second video clips with native synchronized audio in a single pass . It supports text-to-video, image-to-video, video-to-video, keyframe-to-video, and agentic chaining of multiple clips . Output reaches 1080p resolution at 24 fps with multi-shot consistency and camera controls .

Audio generation — Audio is produced natively alongside video, including multilingual dialogue and sound effects that are causally linked to visual events . The output is 48 kHz stereo .

Image generation and editing — The model synthesizes and edits images across diverse styles, aspect ratios, and resolutions, with improved text rendering and complex prompt handling . Early access for image generation was scheduled to follow "in the coming weeks" after launch .

Action prediction — FLUX 3's world understanding extends to predicting robotic actions, either natively or via a finetuned action decoder . This capability turns the model into a shared backbone for both creative content generation and physical robot control .

Unified architecture — Built on BFL's Self-Flow approach, which aligns multimodal generation and representation learning in a single framework . An interesting side-effect: adding action prediction temporarily reduces video quality by up to 10%, but the model recovers full quality after roughly 3,500 training steps while retaining action prediction capabilities .

FLUX-mimic: Real-World Deployment at Audi

BFL partnered with Swiss startup mimic robotics to develop FLUX-mimic, a video-action model built on the FLUX 3 backbone . The system trains a lightweight action decoder on intermediate features from FLUX 3's video prediction path, decoding physical actions from the model's learned world representation .

FLUX-mimic has been tested and deployed on real production tasks at Audi . According to the partners, Audi has used these robots to solve "complex soft body manipulation work that would have been simply impossible" with traditional programming . The system reduces the amount of demonstration data needed to learn new tasks compared to existing approaches . The partnership combines BFL's visual foundation model expertise with mimic's robot learning, dexterous manipulation, and production deployment capabilities .

Performance Benchmarks: What We Know and Don't Know

The specific latency figures sometimes discussed in community forums — under 80 ms inference on a single RTX 5090 GPU and 101 ms system reaction time — do not appear in the available official sources from BFL, Bloomberg, or the official press releases . The official BFL blog posts and press coverage from launch day (July 23, 2026) do not include these hardware-specific benchmarks .

Important caveat: These figures may come from an unreleased technical report, a third-party analysis, or later testing, but there is insufficient evidence in the current source set to verify or cite them. For context, the broader FLUX 3 system is designed for near-real-time robotic control, but no official latency specification on consumer GPUs has been published yet.

For comparison, on the predecessor model FLUX.1, an RTX 5090 generates a 1024×1024 image in roughly 5–12 seconds (depending on optimization), and community benchmarks on FLUX.2-dev also report times in the seconds range per image . FLUX 3's video and action prediction workloads will differ substantially, but the claimed sub-100-millisecond latency for action inference is not yet confirmed by BFL.

What the Model Can and Cannot Do

Verified capabilities

• Generate up to 20-second video with native audio from text, images, or reference video
• Synchronize multilingual dialogue and sound effects to visual events
• Predict robot actions and control dexterous manipulation tasks
• Run on a single consumer GPU for robotic control applications

What is not in the official sources

• No specific latency benchmarks (e.g., inference time on RTX 5090) have been published by BFL
• No technical report with full architecture details or parameter counts is yet available
• Image generation was not available at launch; early access was promised "in the coming weeks"
• Open-weight release ("FLUX 3 Dev") is planned for later but with no firm date

Early Access Availability

• FLUX 3 Video (with audio) — available in early access immediately as of July 23, 2026, via BFL's models page and a waitlist .
• FLUX 3 Image — early access was announced to follow "in the coming weeks" .
• Action prediction (FLUX-mimic) — available through selected research and commercial partners, beginning with mimic robotics .
• Open-weight release ("FLUX 3 Dev") — planned for later, providing a multimodal backbone for both content creation and action prediction .
• Access can be requested via BFL's early access form at bfl.ai/models/flux-3 .