Black Forest Labs (BFL) is increasing its FLUX household past picture technology with at this time's launch of FLUX 3, a multimodal frontier mannequin educated to grasp and generate photos, or mixed audio/video clips as much as 20 seconds from a single immediate — and to increase the identical underlying structure to robotic imaginative and prescient and actions.
The Freiburg, Germany-based AI lab says FLUX 3 is collectively educated throughout these modalities moderately than assembling separate picture, video and audio fashions behind a typical interface.
That distinction is central to the corporate's pitch: BFL desires enterprises to consider inventive technology, simulation, laptop use and robotics as linked functions of a single functionality it calls visible intelligence — fashions, within the firm's phrases, "that may understand, predict, and act throughout bodily and digital environments." This launch marks BFL's first public video technology mannequin.
FLUX 3 will likely be provided by 4 product traces: FLUX 3 Video, FLUX 3 Picture, FLUX 3 Motion and the upcoming, open supply FLUX 3 Dev. FLUX 3 Video, with non-compulsory native audio technology, and FLUX 3 Motion are getting into a gated "Early Entry" program now, to which anybody can apply, however which BFL should approve.
There may be presently no public entry by BFL's utility programming interface (API) or these of companions but, however the firm says FLUX 3 Picture will roll out within the coming weeks, adopted by common availability. The restricted preliminary availability rollout echoes the discharge methods of recent fashions from different frontier labs within the U.S. recently, together with Anthropic and OpenAI, although these had been ostensibly for safety issues and attributable to authorities request.
What the corporate has not introduced is pricing, manufacturing service-level commitments, analysis methodology, pattern sizes, rater counts or any image-model benchmarks in any respect. Enterprise consumers due to this fact can’t but calculate complete price of possession or independently reproduce the video comparisons.
One other large notable omission: FLUX 3 is not launching with downloadable weights at the moment, nor an open supply license. BFL says quicker and open-weight variations will arrive later this yr, and its technical weblog names FLUX 3 Dev as "open-weight entry to a multimodal spine, for content material creation (video, audio and picture) and motion prediction" — a significantly broader dedication than any earlier FLUX Dev launch, all of which lined photos solely.
However it arrives final within the sequence. Builders accustomed to receiving a regionally deployable FLUX variant alongside — or quickly after — a significant mannequin announcement should wait. That delay doesn’t negate the corporate's dedication, however it’s disappointing given the position open weights have performed in FLUX's adoption to this point.
Flux 3 is rated increased than the competitors, however lacking pricing and benchmarking particulars could stop fast enterprise adoption
BFL has printed a number of benchmark comparisons, however they're certified as preliminary — with full benchmark outcomes and methodology to be printed later throughout broader common availability.
In early head-to-head choice testing on 10-second, 720p text-to-video clips with audio, the corporate says FLUX 3 was most well-liked over Luma Ray 3.2 in 93% of comparisons, Runway Gen-4.5 in 77%, Grok Think about Video in 69%, Kling v3 Professional in 60%, Blissful Horse v1 in 59%, Blissful Horse 1.1 in 57%, and each Seedance 2.0 and Google's Gemini Omni Flash in 52%.
One caveat travels with each a kind of figures, and it comes from BFL itself. The chart carrying the outcomes is labeled a "preliminary analysis of an early FLUX 3 candidate" — that means the numbers describe a pre-release checkpoint moderately than the mannequin now getting into early entry. That cuts each methods: the delivery mannequin could carry out higher, however nothing printed at this time measures what clients will truly name.
Luma Ray 3.2 and Runway Gen-4.5, the place FLUX 3 posted 93% and 77%, are the softest comparisons on the record — established merchandise, however not the fashions at the moment setting the tempo in impartial video rankings. These are actual wins, and they’re those least more likely to change an enterprise shortlist.
Seedance 2.0, at 52%, is a statistical coin flip towards a mannequin most Western enterprises can’t at the moment procure. ByteDance indefinitely postponed Seedance 2.0's worldwide rollout after Netflix, Warner Bros., Disney, Paramount and Sony despatched authorized threats over alleged systematic copyright infringement, and that suspension stays in place. Tying a frozen product is neither a robust declare nor a harmful one.
Gemini Omni Flash, additionally at 52%, issues way more. Omni is the closest large-platform analogue to what FLUX 3 is making an attempt — multimodal enter, video and audio-aware creation, conversational enhancing — and by BFL's personal measurement, the 2 are indistinguishable on 10-second text-to-video high quality.
Google's benefit in that matchup is that Omni is usually out there by way of Google's Gemini API for $0.10 per second of generated 720p video, or a 10-second clip for round.
One regional wrinkle issues for a German firm's residence market. Modifying uploaded video is unavailable to Omni Flash customers within the European Financial Space, Switzerland and the UK, although enhancing video the mannequin itself generated is permitted. A European enterprise that wishes to run its current footage by a generative enhancing cross can’t at the moment achieve this on Omni Flash.
Right here's a tough information for enterprises contemplating which video fashions to depend on:
Mannequin | Max single-generation period | Max decision | Key constraints | Worth per 10-second clip (720p) | Worth per 10-second clip (1080p) | Worth per 10-second clip (4K) |
FLUX 3 Video | 20 seconds | Not acknowledged; evaluations run at 720p | Early entry; no printed SLA or pricing | Not introduced | Not introduced | Not introduced |
HappyHorse 1.1 | 15 seconds | 1080p | No 4K; closed weights | Not printed (v1.0 reseller price is ~$1.82) | Not printed (v1.0 reseller price is ~$3.12) | n/a |
Veo 3.1 | Per-second billing | 4K | Helps clip extension; preview | $4.00 | $4.00 | $6.00 |
Veo 3.1 Quick | Per-second billing | 4K | Preview | $1.00 | $1.20 | $3.00 |
Veo 3.1 Lite | Per-second billing | 1080p | No 4K, no clip extension; preview | $0.50 | $0.80 | n/a |
Gemini Omni Flash | 10 seconds (3s minimal) | 720p at 24 FPS | Preview abd no EU entry | $1.00 | n/a | n/a |
One structure for media technology and bodily motion
FLUX 3 builds on Self-Move, BFL's technique for aligning multimodal understanding and technology inside one structure, publicized again in March 2026.
The corporate says it considerably scaled up compute and knowledge to coach throughout video, photos and audio concurrently, and that testing confirmed video technology and motion prediction don’t require separate foundations — the identical structure might be prolonged to motion prediction with out sacrificing what it realized from video.
"We place imaginative and prescient on the middle of our method as a result of it’s the most signal-rich medium of the bodily world. Photos convey construction, photos and video educate spatial relationships, video teaches dynamics, and actions reveal causal relationships. However imaginative and prescient alone will not be the entire image," stated Robin Rombach, co-founder and CEO of BFL, in a pre-release assertion supplied to VentureBeat. "True intelligence means perceiving the world: predicting the way it will change, taking motion, and studying from the outcomes. Joint coaching inside one unified structure is what is going to get us there, as a result of every coaching modality strengthens the others. Audio conveys timing, prosody, and bodily occasions that elude imaginative and prescient. Language conveys targets, abstractions, and directions that pixels can’t simply categorical."
He put the case extra bluntly elsewhere within the announcement: "You may't cheat actuality. A mannequin that solely learns photos can solely generate photos. However the world will not be fabricated from nonetheless frames. It strikes, sounds, modifications, and responds."
BFL says FLUX 3 targets inventive tooling, media, design, e-commerce and bodily AI, supporting video technology with synchronized audio, exact picture enhancing, product and materials consistency throughout movement, multilingual technology and robotic motion prediction. It’s already being examined by Canva, Burda, Magnific (previously Freepik), Krea and Picsart.
For inventive software program corporations, the attraction is consolidation. A single basis may probably assist storyboarding, picture enhancing, product rendering, video variation and localization with out repeatedly translating property and directions between disconnected fashions.
For robotics groups, the potential worth is knowledge effectivity. Fashions that already encode movement, object conduct and bodily change might have much less task-specific robotic coaching than methods ranging from uncooked demonstrations.
What FLUX 3 Video can truly do
The video tier is probably the most concretely specified a part of the launch, and it settles a query that had been circulating as rumor: FLUX 3 generates clips of as much as 20 seconds with audio in a single technology.
Each video output comes with native audio. For comparability, HappyHorse 1.0 tops out at 15 seconds of 1080p with synchronized audio — although BFL has not acknowledged what decision its 20-second clips run at, and its printed evaluations had been carried out at 720p. Nonetheless, a 20-second lengthy clip from a single immediate is among the many longest but achieved, matching OpenAI's discontinued Sora mannequin.
The potential record BFL printed covers:
Textual content-to-video technology.
Picture-to-video technology, both animating from a beginning body or utilizing photos as visible references.
Video-to-video technology from a reference clip, carrying components similar to a selected character into a brand new scene or context.
Generative video-audio continuation from current video and audio enter.
Keyframe-to-video technology for managed transitions between outlined moments.
Multilingual dialogue.
A broad vary of visible kinds and side ratios, from candid camcorder footage to animation and cinematics.
Typography technology and animated design.
Agentic chaining of particular person clips into longer, multi-shot sequences.
That final merchandise is the one enterprise video groups ought to have a look at hardest. BFL claims the capabilities mix to provide sequences lasting a number of minutes, with visible references maintaining characters constant throughout scenes. If that holds up beneath manufacturing circumstances, it addresses the constraint that has saved generative video out of most business pipelines: not clip high quality, however continuity throughout pictures.
It is usually the aptitude the place competitors is most direct. HappyHorse 1.1's headline improve is R2V, or Reference-to-Video, which accepts a number of character reference photos to carry identification secure throughout generated footage — the identical drawback, approached on the enter layer moderately than by agentic clip chaining. Alibaba additionally claims zero-drift lip sync and has particularly focused the artifacts that mark business AI video as artificial, together with facial oiliness and over-sharpening. Character consistency is the place this class is being contested, and each corporations understand it.
BFL says FLUX 3 Video is already significantly sturdy at human facial expressions, associating sounds with bodily occasions, and multilingual output. On the picture facet, the corporate says preliminary evaluations carried out throughout midtraining present important enchancment over earlier FLUX variations in advanced immediate dealing with and textual content technology, together with high-accuracy textual content in a number of languages. It printed no picture benchmarks or win charges.
FLUX-mimic checks whether or not video fashions can turn into robotic fashions
BFL is making use of its unified-architecture thesis by FLUX-mimic, a video-action mannequin constructed on FLUX 3 and developed with Swiss agency Mimic Robotics, one of many first companions to obtain early entry.
The technical weblog describes two distinct routes to motion prediction: integrating native motion prediction instantly into FLUX 3, scaling up the preliminary Self-Move work; and utilizing the pretrained video spine as a dynamics-aware basis from which specialised motion fashions may be finetuned with restricted task-specific knowledge. FLUX-mimic is the second route — the FLUX 3 spine mixed with mimic's robot-learning and production-deployment experience in dexterous manipulation.
FLUX-mimic is designed for general-purpose robotic manipulation: serving to robots perceive a visible scene, predict the implications of an motion, and adapt to new duties with far much less task-specific knowledge.
BFL and Mimic Robotics say that relying on job issue, the mannequin may be finetuned for a selected manipulation job with as little as half-hour of robotic knowledge, the place prior approaches have required 30 or extra hours.
"The toughest a part of robotics is knowledge," stated Elvis Nava, CTO of Mimic Robotics, in an announcement supplied to VentureBeat. "Each new job usually means hours of a robotic repeating itself. As a result of FLUX-mimic is constructed on prime of frontier video fashions that already perceive how the bodily world behaves, it picks up a brand new job in minutes, not days. This manner, we are able to leapfrog the present cutting-edge in robotic studying."
BFL argues {that a} mannequin educated solely on photos can’t perceive a world that "strikes, sounds, modifications, and responds," and that bodily understanding is what produces convincing generated footage. Google makes a virtually similar declare for Gemini Omni.
Its developer documentation cites "world information" that mixes "an understanding of physics" with Gemini's grasp of historical past, science and cultural context. Its advertising is blunter nonetheless: "Most AI fashions simply predict the subsequent pixel to construct a story or a picture. Gemini Omni is completely different," the corporate posted in June, crediting the mannequin with "an intuitive understanding of forces like gravity, kinetic power, and fluid dynamics for extra reasonable actions that comply with real-world logic."
The sensible consequence for enterprise consumers is that world-model language will not be a differentiator. Two of the three main video methods now market bodily understanding as their central benefit, and neither has printed a benchmark that measures it.
There is no such thing as a customary check for whether or not generated water behaves like water, whether or not a dropped object falls at a believable price, or whether or not a sound arrives when the affect does. Human choice scores seize a few of it not directly. Nothing else on provide captures it in any respect.
Open weights helped make FLUX an business customary
BFL formally launched in summer season 2024 and gained a reputation for itself within the AI business within the intervening two years for its dedication to open sourcing high-quality AI picture fashions beloved by builders, creatives, and enterprises.
The corporate's founders, together with Rombach, Andreas Blattmann and Patrick Esser, beforehand helped create VQGAN, latent diffusion and Secure Diffusion, the latter the open supply expertise that kicked off broad AI technology capabilities for the lots and at the moment utilized by many AI picture mills and firms.
That attain translated into business distribution. FLUX fashions now energy generative options inside Adobe Photoshop, Picsart and Nous Analysis's Hermes Agent, amongst different platforms, and the corporate cites movie director Martin Scorsese amongst skilled customers.
Wired journal described Black Forest Labs as a comparatively small firm that however grew to become a number one competitor to Silicon Valley's largest AI labs, with FLUX fashions rating close to the highest of picture benchmarks and turning into a few of the most downloaded text-to-image fashions on AI code sharing neighborhood Hugging Face. The corporate says it now runs a 100-person group throughout Freiburg and San Francisco.
FLUX.1 Dev, FLUX.1 Kontext Dev, FLUX.1 Fill Dev and associated management fashions, launched shortly after the agency's launch, gave researchers and creative-tool builders entry to downloadable checkpoints, native inference and integrations with frameworks together with Hugging Face Diffusers and ComfyUI. FLUX.1 Kontext Dev, for instance, was launched as an open-weight mannequin for analysis and noncommercial use, with generated outputs permitted for business functions beneath the relevant license.
The corporate continued that sample with FLUX.2 Dev in late 2025, a 32-billion-parameter open-weight mannequin combining technology and multi-reference enhancing. Black Forest Labs referred to as it the strongest open-weight picture technology and enhancing mannequin out there at launch and launched weights, reference inference code and optimized implementations for client Nvidia GPUs.
FLUX 3 Dev raises the stakes on that analysis. Earlier Dev releases had been picture fashions. This one is described as a multimodal spine spanning video, audio, picture and motion prediction — that means a single license will govern whether or not an organization can regionally deploy a mannequin that touches each content material manufacturing and bodily equipment. BFL hasn't but shared details about its license, the parameter rely, quantizations or {hardware} necessities.
The corporate frames open weights as an enterprise function moderately than a neighborhood gesture, arguing they permit safe, low-latency native deployment for functions like robotic management methods and let groups adapt FLUX 3 to their very own knowledge, merchandise and workflows.
The monetary backing behind FLUX 3 is value noting alongside the technical claims. Black Forest Labs is valued at $3.25 billion and has raised greater than $450 million from traders together with a16z, AMP, Salesforce Ventures, Nvidia, Common Catalyst, Adobe Ventures, Figma Ventures, Canva and Deutsche Telekom's T.Capital.
