Black Forest Labs releases FLUX 3: A multimodal flow model for image, video, audio and robot action prediction. July 26, 2026 Black Forest Labs (BFL) has released FLUX 3, a multimodal foundation model that learns from images, videos and audio inside a single architecture. It is also the first FLUX model to ship video, audio and action prediction from one set of weights. The Black Forest Labs (BFL) research team argues that no single modality gives a complete description of the world. Images capture spatial structure at one instant. Video restores time and exposes physical dynamics. Audio reveals causal relationships between mechanical events and sound. Each is treated as a lossy projection of the same underlying reality. Training on all of them at once means the modalities constrain each other. The sound has to match the impact. The motion has to obey the mass. The research team calls FLUX 3 its first model built entirely on that principle. The method underneath: Self-Flow. FLUX 3 builds on Self-Flow, BFL's method for aligning multimodal generation and understanding in one architecture. Self-Flow combines the flow matching objective with a self-supervised feature reconstruction objective. The reference implementation on GitHub is Apache-2.0 and uses SiT-XL/2 with per-token timestep conditioning. It trains with a 25% per-token mask ratio and self-distillation from an EMA teacher at layer 20 to a student at layer 8. That released checkpoint is an ImageNet 256x256 research model, not FLUX 3. BFL states that it 'significantly scaled up compute and data resources' on the same approach to train FLUX 3 across video, images and audio simultaneously. Self-Flow itself was introduced in March 2026, so it is not new to this launch. What is new is the scale. What FLUX 3 Video does. FLUX 3 Video generates clips up to 20 seconds long in a single generation, with native audio. The supported modes cover text-to-video, image-to-video, video-to-video from a reference clip, keyframe-to-video for controlled transitions, and generative video-audio continuation from input video and audio. BFL also lists multilingual dialogue, agentic chaining of clips into multi-shot sequences, and strong typography generation with animated designs. The BFL team reports particular strength in human facial expressions and in associating sounds with physical events. Performance. BFL team published preliminary human preference results. The setup was 10-second text-to-video clips at 720p with audio. FLUX 3 was preferred over Luma Ray 3.2 in 93% of comparisons and over Runway Gen-4.5 in 77%. Against Grok Imagine Video the figure is up to 69%, then Kling v3 Pro at 60%, Happy Horse v1 at 59% and Happy Horse 1.1 at 57%. Against Seedance 2.0 and Gemini Omni Flash the result is 52%, close to a coin flip. Interactive explorer. bfl@ flux-3 :~/real-world-models Early Access flux3 $ flux3 -spec One backbone, four output types. * Model class - Multimodal flow matching foundation model * Joint training modalities - image + video + audio, one architecture * Extended output - Action prediction (robot state) * Max video length - 20 seconds in a single generation * Audio - Native, generated jointly with video * Video modes - text-to-video, image-to-video, video-to-video, keyframe-to-video, video+audio continuation * Other capabilities - multilingual dialogue, agentic multi-shot chaining, typography and animated design * Announced - 23 July 2026, Black Forest Labs * Pricing - not published BFL describes each modality as a lossy projection of one underlying reality. Training on all of them at once lets their mutual constraints supervise each other. flux3 $ enter

Anthropic debuts Claude Opus 5, Big Tech rallies for open AI, DARPA flies autonomous F-16 - AI news briefing. Top 7 stories. 1. Anthropic launches Claude Opus 5. Anthropic unveiled Claude Opus 5, the latest iteration of its flagship model and what the company calls its most capable and reliable AI system yet. The new model delivers substantial improvements across reasoning, coding, mathematics, and long-context tasks, with Anthropic claiming state-of-the-art performance on several key benchmarks. Early reports indicate Opus 5 narrows - and in some areas surpasses - the gap with competing frontier models from OpenAI and Google. The release comes at a critical moment as Anthropic scales infrastructure through its AMD partnership and positions Claude as the enterprise-safe alternative in an increasingly crowded market. 2. Nvidia, Microsoft, and Meta lead push against U.S. Open-Weight AI restrictions. A coalition of tech heavyweights led by Nvidia, Microsoft, and Meta is urging the U.S. government to reject proposed restrictions on open-weight AI models, warning that overregulation would cede American leadership to China and stifle innovation. In a coordinated lobbying effort reported by CNBC, the companies argue that open-weight models - whose parameters are publicly available - are essential for research, transparency, and the broader AI ecosystem. The pushback comes as policymakers draft export controls and domestic regulations aimed at preventing advanced AI from falling into the hands of adversaries. The debate has become one of the defining AI policy battles of 2026, pitting national security hawks against an industry that sees openness as a strategic necessity. 3. DARPA and U.S. Air Force successfully fly ai-controlled F-16. In a milestone for autonomous military aviation, DARPA and the U.S. Air Force announced the successful flight of an AI-controlled F-16 fighter jet, demonstrating that machine learning systems can handle complex, high-stakes aerial maneuvers without human intervention. The test, conducted under DARPA's Air Combat Evolution (ACE) program, involved an AI agent piloting the aircraft through a series of combat-relevant scenarios, including evasive maneuvers and target engagement. While the Pentagon emphasizes that human oversight remains in the loop, the demonstration signals a future where AI wingmen fly alongside human pilots - and raises urgent ethical and strategic questions about autonomous weapons. 4. OpenAI GPT-5.6 models arrive on Amazon Bedrock. OpenAI's latest GPT-5.6 family - including Sol, Terra, and Luna variants - is now generally available on Amazon Bedrock, AWS's managed AI service. The move represents a significant expansion of OpenAI's distribution strategy beyond its own platform and Microsoft Azure, giving AWS customers direct access to frontier models through their existing cloud infrastructure. The three model tiers offer different price-performance profiles: Sol for maximum capability, Terra as the balanced workhorse, and Luna optimized for cost-sensitive, high-throughput workloads. The Bedrock integration intensifies the three-way cloud AI platform war between AWS, Microsoft Azure, and Google Cloud, each racing to be the default gateway to the latest models. 5. Oracle fires 21,000 employees to fund AI spending. Oracle laid off approximately 21,000 employees this week in what the company described as a restructuring to redirect capital toward AI infrastructure and product development. The cuts, among the largest in tech industry history, underscore the brutal calculus facing legacy enterprise companies: either retool aggressively for the AI era or risk irrelevance. Oracle has been investing heavily in AI-optimized cloud data centers and plans to deploy tens of thousands of NVIDIA GPUs. The layoffs have drawn sharp criticism from labor advocates and renewed debate about whether the AI productivity boom will create or destroy more jobs in the near term. 6. Alphabet's soaring AI capex rattles investors. Alphabet's latest earnings revealed a dramatic surge in capital expenditures driven almost entirely by AI infrastructure spending, sending a chill through Wall Street and raising broader alarm about Big Tech's AI spending trajectory. Reuters reported that the company's cash burn has accelerated faster than analysts anticipated, with no clear timeline for when these investments will translate into proportional revenue growth. The reaction mirrors concerns across the sector: Microsoft, Meta, and Amazon are all pouring tens of billions into AI compute, and investors are growing impatient for returns. Alphabet's results may be a bellwether for whether the AI capex super-cycle is sustainable - or a bubble in the making. 7. Black Forest Labs releases Flux 3 Mimic video-action model. Black Forest Labs launched Flux 3 Mimic, a new generation of video-action models that can translate text descriptions and reference images into realistic, temporally coherent video sequences with unprecedented motion fidelity. The model builds on the Flux family's reputation for photorealism and adds sophisticated action understanding, enabling outputs ranging from subtle character animations to complex physical interactions. The release heats up the already competitive AI video generation market, where players like Runway, Pika, and OpenAI's Sora are racing to deliver production-quality output. Flux 3 Mimic is available via API and through Black Forest's creative tooling platform. Trend watch. | Story | Impact | Why It Matters | | Claude Opus 5 release | Raises frontier model bar | Anthropic is proving it can sustain a multi-front competition - model quality, enterprise trust, and infrastructure scale - against OpenAI and Google simultaneously. | | Open-weight regulation fight | Defines U.S. AI policy | The outcome will determine whether cutting-edge AI remains accessible to researchers and startups or becomes locked behind corporate and government gates. | | DARPA AI F-16 flight | Military AI goes operational | Autonomous combat systems are no longer theoretical. This test forces overdue conversations about AI in lethal decision-making. | | Oracle's 21K layoffs | AI-driven workforce disruption | When a company fires 21,000 people to fund AI, it crystallizes the tension between AI-driven efficiency and human employment. | | Alphabet's AI capex surge | Tests investor patience | If Alphabet can't show returns, the AI spending spree across Big Tech could face a reckoning - with ripple effects for the entire AI supply chain. | What to watch. Claude Opus 5 vs. GPT-5.6 benchmarking wars. Independent evaluators will spend the coming weeks stress-testing Opus 5 against OpenAI's latest models. Expect a flood of benchmark comparisons, leaderboard updates, and hot takes - with real enterprise procurement decisions hanging in the balance. Open-weight regulation reaches a decision point. The Nvidia-Microsoft-Meta lobbying blitz, combined with startup founder pressure, is forcing the administration toward a resolution. A draft executive order or legislative text could drop within weeks, and its scope will signal whether the U.S. embraces or restricts open AI. Military AI ethics enters the mainstream. The DARPA F-16 flight will almost certainly trigger congressional hearings and renewed calls for international agreements on autonomous weapons. Watch for the Pentagon to release detailed safety frameworks in an attempt to get ahead of the narrative. Big Tech earnings spotlight AI ROI. With Alphabet's cash burn in the headlines, upcoming earnings from Microsoft, Amazon, and Meta will be scrutinized through the same lens. Any sign of slowing AI investment - or conversely, doubling down without clear returns - will move markets.

Black Forest Labs unveils FLUX 3: an image lab bets its future on multimodal frontier models. On July 23, 2026, Black Forest Labs announced FLUX 3, a single model trained jointly on images, video and audio and extensible to actions. A neutral, analytical look at what the German lab is claiming, what has actually shipped, and why an open-weight image house is pivoting to visual intelligence and physical AI. Black Forest Labs is the German company that made its name in 2024 with FLUX.1, a set of open-weight image models that quickly became a default for developers who wanted to generate pictures without renting a closed API. On July 23, 2026, it announced something considerably more ambitious: FLUX 3, which the lab describes not as a better image model but as a multimodal frontier model for visual intelligence, trained jointly on images, video and audio in one architecture and extensible to predicting actions. It is a large bet, and like most large bets it is worth separating from the marketing. This piece looks at what FLUX 3 is claimed to be, what has actually shipped, and why a still-image lab is reframing itself around video, audio and robotics. From still images to one shared model. The core technical claim is architectural. Rather than bolting a video generator and an audio generator onto an image model, FLUX 3 is trained across modalities at the same time, so that images, video and audio are learned within a single set of weights. Black Forest Labs credits an approach it calls Self-Flow, which it describes as a way to align multimodal generation and understanding inside the same underlying architecture, then scaled up with more compute and data. The lab also says the same architecture can be extended to predict actions, which is the bridge from generating media to controlling machines. From still images to visual intelligence. Black Forest Labs' public model lineage, showing the jump from single-modality image models to a unified multimodal system. * 2024 FLUX.1 launches with the company Open-weight text-to-image models put the new German lab on the map, with weights anyone could download and run. * 2025 FLUX.1 Kontext adds in-context editing The line moves from generating images to editing them from instructions, keeping the open-weight release model. * Jul 23, 2026 FLUX 3 goes multimodal One architecture jointly trained on images, video and audio, extensible to actions. The lab reframes itself around "visual intelligence" rather than still images. The strategic logic is easier to state than the engineering. An image-only lab competes in a crowded, fast-commoditising market where open weights and falling prices squeeze margins every quarter. A lab that owns a unified model for image, video and audio, and can point it at robotics, is playing for a much larger prize and a more defensible position. Whether the single-architecture approach actually produces better results than specialised models is the open technical question, and it is not one an announcement can settle. What has actually shipped, and what has not. This is where a careful reading matters, because the gap between the announcement and the available product is unusually wide. One model, four modalities. What FLUX 3 is designed to generate, and how far along each part was at announcement. Video, audio and action lead; the image release trails. Image Not yet shipped The lab's original domain, now one head of a shared model rather than a standalone system. Expected "in the coming weeks" Video In early access Text-to-video up to 20 seconds, carrying native audio that is generated in sync rather than added afterward. Early access via API Audio In early access Learned jointly with image and video, so sound and picture come from the same underlying representation. Part of the video model Action In early access Predicting actions rather than pixels, the basis of a robotics variant the lab says it is testing on production lines. Early access, select partners The headline capability, text-to-video of up to 20 seconds with native audio generated in sync rather than dubbed on afterward, is in early access through an API, with private weights shared to a set of selected partners. The action and robotics side is likewise early access to a small group. But FLUX 3 Image, the part closest to the lab's existing business and the thing most of its users actually want, is described only as expected in the coming weeks. The fully open-weight release, FLUX 3 Dev, is planned for later in 2026. In other words, the lab announced a multimodal frontier model whose most in-demand component and whose open weights are both still to come. None of that makes the claims false, but it does change how they should be read. An announcement with a video demo and a coming-soon image model is a statement of direction and a bid for attention and partners, not a shipped product a developer can benchmark today. The honest position is that FLUX 3's capabilities are, for now, mostly demonstrated rather than generally available, and the real test is whether the shipped models match the reel when the weights arrive. The physical AI turn. The most striking part of the announcement is the least expected from an image lab. Black Forest Labs says FLUX 3 marks its first move into physical AI, with a robotics variant already being tested on Audi production lines. The connective idea is that a model which has learned how the visual world moves, by training on video, has learned something reusable about physics and cause and effect, and that predicting a robot's next action is not so different from predicting the next frame of a video. This puts Black Forest Labs on the same road that several larger players are already travelling, from Nvidia's world-model work to the embodied-AI efforts at the big labs. The company arrives with a real asset, a strong track record in visual generation, and a real disadvantage, none of the robotics deployment experience the incumbents have. A test on an automotive line is a meaningful signal of intent and access, but it is a pilot, not a product, and physical AI has a long history of impressive demos that struggle to generalise beyond the exact station they were tuned for. Why it matters beyond one lab. Strip away the specifics and FLUX 3 is a clean example of two patterns shaping the year. The first is convergence: separate model categories, image generation, video, audio, and now action, are collapsing into single multimodal systems, because the same underlying representation of the visual world can in principle serve all of them. The second is the pressure that convergence puts on specialists. A lab that does one modality well is exposed if the general multimodal models become good enough, so the rational move, even for a company as identified with images as this one, is to widen out before the ground shifts underneath it. For anyone building products on top of generative media, the takeaway is not to pick a winner today. It is that the set of usable models for any given visual or multimodal task is going to keep expanding and reshuffling, as image specialists move into video, video labs move into audio, and general labs move into all of it. The teams that benefit are the ones set up to swap in whichever model is best for a given job as the field churns, rather than wiring a pipeline to a single provider's current lineup. A model-agnostic workspace such as Metir AI reflects the same principle on the language side: treat the model as a component you can route around, not a foundation you are locked to. The bigger picture. FLUX 3 is best understood as a well-executed announcement of intent from a lab that has earned the right to be taken seriously in visual generation and is now reaching well beyond it. The multimodal architecture, the in-sync audio, and the robotics pilot are all genuine and all early. The image model everyone actually wants, and the open weights that made the company's name, are still ahead. That combination, real ambition paired with a product that is mostly still in early access, is the fair summary. The reel is impressive; the meaningful verdict waits for the weights. Sources: Header image: the Schwabentor gate tower in Freiburg im Breisgau, the German city on the edge of the Black Forest where Black Forest Labs is based, by Jorge Franganillo via Wikimedia Commons, licensed under CC BY 2.0. In-body photograph of a KUKA industrial robotic arm executing an artwork by Oleg Yunakov via Wikimedia Commons, licensed under CC BY-SA 4.0. The arm shown illustrates the action modality and is not FLUX 3 hardware.