- Flux 3 can generate video with synchronized audio for up to 20 seconds from text, images or existing video, according to Black Forest Labs.
[[1]](https://bfl.ai/blog/flux-3) - Access requires an application; Black Forest Labs has not published Flux 3 pricing, downloadable weights or a release-specific technical report.
[[1]](https://bfl.ai/blog/flux-3)[[2]](https://bfl.ai/models/flux-3) - The company’s benchmark and robotics results have not yet been independently reproduced.
[[3]](https://the-decoder.com/flux-3-generates-videos-with-native-audio-up-to-20-seconds-long-a-first-for-black-forest-labs/)[[4]](https://www.mimicrobotics.com/blog/introducing-flux-mimic)
Black Forest Labs has opened gated early access to Flux 3, its first model trained jointly across images, video and audio. The system can produce video with native sound for up to 20 seconds in one generation and supports text-to-video, image animation, video transformation, keyframe transitions and multilingual dialogue.[1]
The launch is narrower than a general release. Prospective users must request access, while the company’s public pricing page and API documentation do not list Flux 3. Black Forest Labs says video and audio tools will eventually arrive through APIs and private weights, with an open-weight Flux 3 Dev backbone planned later.[1][2]
Strong company scores, limited outside evidence #
In Black Forest Labs’ preliminary evaluation of 10-second, 720p clips, viewers preferred Flux 3 over Runway Gen-4.5 in 77% of comparisons and over Luma Ray 3.2 in 93%. Results were much closer against Seedance 2.0 and Gemini Omni Flash, with Flux 3 receiving 52% preference against each. The company said its evaluation harness remains under development; The Decoder reported that independent tests were not yet available.[1][3]
Flux 3 builds on Self-Flow, a method described in a March research paper that combines representation learning with generative flow matching. That paper covers the underlying training approach, but Black Forest Labs has not released a Flux 3 technical report detailing model size, training compute, full datasets or safety evaluations.[1][5]
Robotics claims come from the developers #
Black Forest Labs and mimic robotics also introduced Flux-mimic, which adds an action decoder to Flux 3’s video backbone. Mimic reported 95% success on a soft-body kitting task, compared with 55% for its adapted pi0.5 model, and said Audi is testing and deploying the system on factory tasks. Those figures come from the partner developing the system and have not been independently audited.[4]
Black Forest Labs says the phased rollout will include safety testing. Its broader usage policy prohibits unlawful deepfakes, election disinformation, non-consensual explicit content and military or surveillance applications, while its responsible-development policy describes filtering training data and moderating prompts, uploads and outputs where legally required. The company has not published Flux 3-specific red-team results or a system card.[6][7]
Further sources #
[[1] Black Forest Labs, “FLUX 3 - Real World Models: Towards Multimodal Flow Models … ↗](https://bfl.ai/blog/flux-3)
[[2] Black Forest Labs, FLUX 3 model and pricing pages, accessed July 24, 2026. ↗](https://bfl.ai/models/flux-3)
[3] The Decoder, “Flux 3 generates videos with native audio up to 20 seconds long, … ↗
[[4] mimic robotics, “Introducing FLUX-mimic: Scaling Video-Action Models for Genera… ↗](https://www.mimicrobotics.com/blog/introducing-flux-mimic)
[[5] Chefer et al., “Self-Supervised Flow Matching for Scalable Multi-Modal Synthesi… ↗](https://arxiv.org/abs/2603.06507)
[[6] Black Forest Labs, Usage Policy, revised July 14, 2026. ↗](https://bfl.ai/legal/usage-policy)+1 more
The stories that matter, in one email. Free — unsubscribe anytime.