Ignore the 20-second clips for a minute. The most interesting thing about FLUX 3, which Black Forest Labs launched on July 23, is a word that appears in the singular: model. One flow-matching architecture, trained jointly on images, video, and audio, that also predicts robot actions. A year ago that sentence would have described 4 separate products.
The same 3 weeks delivered the theory to match. A paper from some of the most cited names in computer vision showed that video generation backbones beat specialist vision models at perception tasks with a fraction of the training data. Adobe researchers demonstrated a video model you can move a camera through after the footage exists. Google put a price on conversational video: 10 cents a second. Each of these is a separate story. Read together, they say the tool most of us call a video generator is a temporary form factor, and the labs already know it.
- FLUX 3 is one set of weights for image, video, audio, and robot action. Video ships first; the open FLUX 3 Dev backbone ships last, later in 2026.
- New research (arXiv 2607.09024) shows video generation pretraining transfers to depth, segmentation, and pose with 7x to 500x less task data than specialist models.
- Gemini Omni Flash prices generated video at $0.10 per second and lets you revise a clip by talking to it.
- Open weights are becoming the trailing tier of frontier multimodal releases. Kimi K3 is the giant exception, with strings attached.
One set of weights, 4 jobs
FLUX 3 generates video up to 20 seconds with native, jointly generated audio: the soundtrack comes out of the same forward pass as the pixels, multilingual dialogue included. It handles text-to-video, image-to-video, video-to-video, and keyframe conditioning (pin the frames you need and let the model fill the motion between them). In Black Forest Labs’ own preliminary human preference tests at 10 seconds and 720p, it wins 77% of matchups against Runway Gen-4.5 and 93% against Luma Ray 3.2. It also ties Seedance 2.0 and Gemini Omni Flash at 52%.
The numbers are self-reported and should be treated that way. The architecture is the real story. BFL did not bolt an audio model onto a video model. It trained one backbone across modalities and is now peeling off products from it in sequence: FLUX 3 Video first, in application-gated early access. Then FLUX 3 Action, the robotics head, already being tested on production tasks at Audi with mimic robotics. Then FLUX 3 Image. Last, and promised only for “later in 2026,” FLUX 3 Dev, the open-weight multimodal backbone. Parameter count and pricing remain undisclosed.
The research got there first
Video generation models are already general-purpose vision models, and as of this month that is a measured result. On July 10, 2 weeks before the FLUX 3 launch, a paper landed with a title that reads like a thesis statement: “Video Generation Models are General-Purpose Vision Learners” (arXiv 2607.09024). The author list includes Andrew Zisserman, Joao Carreira, and Kaiming He. The idea: take a text-to-video diffusion backbone and, instead of asking it for clips, instruct it through text to produce depth maps, surface normals, camera pose, segmentation masks, and 3D keypoints.
Their model, GenCeption, matches or beats specialist systems including DepthAnything3 and SAM3, and against D4RT and VGGT-Omega it reaches comparable performance on 7x to 500x less training data. It generalizes from synthetic humans to real animals and robots. That is the empirical case that video generation pretraining is to vision what language model pretraining was to NLP. To predict the next frames of the world, a model has to learn depth, occlusion, physics, and object permanence. Those representations were sitting inside the clip generators all along. The model you rent to make b-roll is quietly one of the best general-purpose eyes money can buy, and every frontier lab can read that paper too.
Two more July papers push the same direction from the world-model side. Wonder (arXiv 2607.26037, Johns Hopkins and Adobe Research, July 28) is an interactive video world model that holds a scene coherent at minute scale and 16 frames per second, and its sharpest trick is video-conditioned generation: hand it existing footage of a dynamic scene and re-shoot that scene in real time along a camera path that was never filmed. AlayaWorld (arXiv 2607.06291, July 7) open-sources the full stack for playable generative worlds, from data pipeline to deployment, where the next frame is computed from world state plus your action. One is a filmmaking primitive: coverage you never shot. The other is a reproducible baseline anyone can extend. Both treat video generation as an interface to a simulated world rather than a rendering endpoint.
Revision replaces re-rolling
Google’s Gemini Omni Flash, which opened to developers at the end of June and rolled into Google Vids in July, completes the picture from the interface side. Any mix of text, image, audio, and video in; video with audio out; clips capped at 10 seconds in preview, priced at $0.10 per second of output. The differentiator is statefulness: you edit a generated clip by giving notes in conversation instead of re-prompting from zero.
Anyone who has produced with current video models knows the slot-machine loop: generate 20 candidates, cherry-pick, burn budget. A stateful session shifts the cost model from per-attempt to per-refinement, and it shifts the skill from prompt incantation to direction. “Same shot, warmer light, hold on her hands” is a sentence a director says. The first API workflows where that sentence works are here, and at roughly $1 per 10-second clip the economics of iteration change too.
The open-weights squeeze
Open weights are no longer part of a frontier launch. They are a later tier of it. Look at what shipped weights this month and when.
| Release | Weights at launch | The catch |
|---|---|---|
| FLUX.1 dev (2024) | Yes, day one | Non-commercial license. Became the backbone of the open image stack, from ComfyUI pipelines to the LoRA fine-tune ecosystem. |
| Qwen-Image 1.0 / 2.0 | Yes, day one | Apache 2.0. The permissive baseline everyone standardized on. |
| Qwen-Image-3.0 (Jul 21) | No | Invite-only API. No benchmarks, no technical report, no license, no promise of weights. |
| FLUX 3 (Jul 23) | No | API and gated early access first. Open FLUX 3 Dev backbone promised for “later in 2026,” last in the rollout queue. |
| Kimi K3 (Jul 27) | Yes, day one | Kimi K3 License: MIT text plus conditions. Resellers above $20M over any 12 months need a separate agreement; products over 100M MAU or $20M monthly revenue must display “Kimi K3” in the UI. |
Qwen-Image-3.0 is the starkest data point. Alibaba’s image line shipped permissive weights on day one for 2 generations. The third generation, which renders full infographic grids from 4.5K-token prompts with legible text down to roughly 10 pixels, arrived as a closed, invite-only API with no paper and no stated plan to open. Teams that standardized on Apache-licensed Qwen-Image now face an upgrade path that runs through a closed endpoint.
Kimi K3 cuts the other way, and the exception is instructive. Moonshot released full weights for the largest open model to date: 2.8T total parameters in a sparse mixture-of-experts (16 of 896 experts active per token), a 1M-token context window, and native image and video understanding, quantized to MXFP4 so it is actually downloadable. API pricing is aggressive: $0.30 per million input tokens on cache hits, $15 per million output. But even here the license grew teeth. The Kimi K3 License reads like MIT for most of its length, stays free for internal use at any scale, then meters the companies that resell it. Weights are still shipping; unconditional weights are getting rare. For frontier generation in particular, the open tier now trails the API tier by months, and the FLUX.1 precedent suggests it will trail in capability as well.
If you build
- Put an abstraction layer between your product and any one video API. The category is consolidating into everything-models, pricing is in flux, and this month proved that even historically open vendors change posture between versions. Being 2 weeks from a provider swap is a feature you build now.
- Do not plan 2026 around fine-tuning a frontier open video model. FLUX 3 Dev is explicitly last in the queue. If your roadmap needs open weights, Kimi K3’s terms make frontier-class understanding self-hostable today; treat open frontier generation as unavailable until proven otherwise.
- Try a video backbone for perception before you reach for a specialist model. The GenCeption result says the depth, segmentation, and pose you need may already live in models you can rent, at a fraction of the labeled data. That is a cost lever most stacks have not priced in.
If you create
- Learn to direct in sessions. Conversational revision rewards the vocabulary of a director working with a cinematographer. That craft transfers straight from set to model, and it is exactly the skill the slot-machine workflow never exercised.
- Watch the re-shoot primitive. Wonder-style video-conditioned generation means footage becomes a scene you can re-cover, and coverage becomes a post-production decision. Blocking and lens choice stay decisions; when they get made is what changes.
- Native audio collapses a pipeline stage. A 20-second take with synchronized dialogue from one forward pass removes the video-then-audio round trip for previz and short-form work. Budget your test time accordingly.
- Image work is inside the same model now. FLUX 3 Image comes off the backbone that generates the video and the audio, and Qwen-Image-3.0 renders full layout grids with legible type from a single prompt. Designers keeping a separate tool for stills are about to be choosing a model family instead.
A prediction you can check us on
Convergence stories are cheap when nothing is falsifiable, so here is a claim with a date on it. When FLUX 3 Dev ships later in 2026, it will land materially behind the API version in capability, the same pattern FLUX.1-dev set in 2024. If BFL proves us wrong and ships the full backbone open, that would be the best news of the year for the open stack, and we will say so in this space. Either way, the standalone video generator you are integrating today is a transitional form. The models are becoming general engines of the visual world. Build and shoot like the merge is already scheduled.