Black Forest Labs Launches FLUX 3: One Model for Images, Video, Audio, and Robot Actions

Black Forest Labs (BFL) released FLUX 3 on July 23 — a multimodal foundation model that handles image generation, video synthesis, audio generation, and robot action prediction within a single architecture, officially jumping from "text-to-image tool" into the omni-model arena.

Three Key Takeaways

Four modalities share one set of weights. FLUX 3's core pitch: images, video, audio, and action prediction all run inside the same flow-matching framework. This is not four models stitched together — it is one model that learned four things at once. Audio is natively synchronized during video generation, meaning visuals and sound come out of the same inference pass with no post-hoc alignment needed.

Staged rollout, open weights last. FLUX 3 Video is already in early access at bfl.ai. FLUX 3 Image and FLUX 3 Action (robot action prediction) follow in the coming weeks. The final tier is FLUX 3 Dev — the open-weight release of the full multimodal backbone, expected later in 2026. BFL's playbook is clear: monetize via API first, then trade open weights for ecosystem.

Going head-to-head with OpenAI and Google. This launch repositions BFL from "the FLUX text-to-image company" to "omni-model contender," directly competing with OpenAI, Google DeepMind, and ByteDance. The differentiator is BFL's commitment to eventually open-source the full weights — a meaningful signal for independent developers and the research community.

WangDou's Take

BFL's move is clever and dangerous in equal measure. Clever because: if you only do text-to-image, you are a Midjourney competitor with a visible ceiling. If you do omni-modal, you are an OpenAI competitor with a much bigger story to tell investors. Dangerous because: one model doing four things has to beat specialists at each one. FLUX 2 genuinely held its own in image generation, but video and audio are already crowded with Sora, Veo, and Kling — rivals that have burned billions. "Unified architecture" sounds elegant, but if your video hits only 70% of Sora's quality, users will not care that it came from the same weights. The real test arrives the day FLUX 3 Dev goes open-source — the community will deliver a verdict more honest than any benchmark within 48 hours.

Source: GlobeNewsWire, MarkTechPost, kie.ai