Frontier video went open-weight in 10 days.

The distance between the best closed video model and the best open one is 5 Elo points. On Artificial Analysis’s text-to-video arena, as read on August 4, Google’s Gemini Omni Flash sat around 1242. MiniMax’s H3, whose weights you can download from Hugging Face, sat around 1237. At that gap, blind human voters prefer the closed model 50.7% of the time. A coin flip, against a model you can run yourself.

Two weeks ago we wrote that open weights were becoming the trailing tier of frontier multimodal releases, and July’s evidence backed it: Qwen-Image-3.0 closed, FLUX 3 Dev queued last. Then 10 days of August argued back. MiniMax posted frontier-class weights on August 3 with day-one ComfyUI support. Lightricks shipped LTX-2.5 open on August 11, tuned for hardware you own. And OpenAI’s exit from video generation entered its final phase: the Sora 2 API shuts down September 24. The open tier did not fall behind the frontier. In video, this month, it caught the frontier while the biggest lab in the world walked away from the category.

10 days, 2 weight drops, 1 funeral date

How to cite: The Multimodal Society, “The 10 Days AI Video Went Open-Weight,” August 2026. multimodalsociety.com/blog/video-went-open-weight

The anchor release is MiniMax H3, the model behind Hailuo 3.0. The platform launched July 31; the weights landed on Hugging Face August 3. It is a 33B-parameter omni transformer: text, image, video, and audio in, video with native 32 kHz stereo audio out of the same forward pass, 4 to 15 seconds per clip, up to 2K at 24 fps. The release is 2 checkpoints. FL2VA covers text-to-video plus first-and-last-frame keyframing. Ref2VA is reference-driven generation, accepting up to 9 reference images plus 3 video and 3 audio clips per generation, per MiniMax’s documentation. ComfyUI merged native support the same day: 4 nodes, 6 workflow templates. The repo passed 83k downloads in its first month.

LTX-2.5, announced August 11, extends Lightricks’ run-it-yourself position. Open weights, day-one ComfyUI support, and native multishot generation: one prompt produces a multi-cut sequence that holds character and scene across the cuts, which pushes edit structure inside the generation itself. A new diffusion video decoder targets the high-motion artifacts that made earlier LTX output easy to spot. Lightricks claims a 10-second clip generates in 6.8 seconds on 2 GB200s; treat vendor speed claims as marketing until benchmarked, but the optimization target is the point. This model is tuned for RTX cards and DGX Spark, and runs on a Mac.

Black Forest Labs sits between the tiers. FLUX 3 Video went GA on August 4, API only: clips up to 20 seconds, native audio with lip-synced multilingual dialogue, keyframe control, and multi-scene generation. The open FLUX 3 Dev backbone we wrote about in July remains exactly where it was: promised, undated, unlicensed, parameter count undisclosed.

3 arenas, 3 different leaders

The Elo chart everyone screenshots is text-to-video. It is 1 of 3 arenas, and each has a different name at the top.

Read the spreads, not the ranks. A 5-point Elo gap converts to a 50.7% win rate, and the image-to-video podium spans 7 points across 3 models from 3 companies on 2 continents. These differences are inside the noise of the voting itself. The honest summary is that text-to-video has a closed leader by a hair, image-to-video belongs to ByteDance’s Seedance 2.0, and the editing arena belongs to an open-weight model. “Which video model is best” now has a different answer per task, and whatever answer you cache goes stale in a week.

Runway read this before the rest of us. Its Media Router, launched July 23, has developers declare cost, latency, and quality preferences instead of naming a model, then routes each call across Runway’s own models and third-party ones, Seedance and ElevenLabs included, with USD price caps per second of output and a dry-run mode that validates routing before spend. When a frontier video lab’s flagship developer product is a router over its competitors, single-model loyalty is over as an architecture and as a skill.

“Open weights” is a license, so read it

Every downloadable release this month ships with conditions. None is open source by any OSI reading, and the conditions differ in ways that decide whether you can ship.

The H3 regional clause is the one to sit with. A model free to use in Jakarta or São Paulo requires a filed application in San Francisco or Berlin. That is a new shape of license fragmentation: geography as a tier. Kimi K3 metered resellers by revenue in July; H3 meters users by jurisdiction in August. The lesson from both is the same. License review is now an engineering step with the same weight as an eval, because “the weights are on Hugging Face” tells you nothing about whether your product can legally exist.

The counter-beat: OpenAI’s exit goes final

While the open tier caught up, the largest AI lab completed its retreat from video generation. Per OpenAI’s own help center, the Sora app and web experience ended April 26. The second stage lands September 24, when the Sora 2 API is removed and remaining user data is deleted after the deadline. That is 6 weeks from today. If Sora 2 is in your production path, the migration is not a roadmap item, it is an incident. Export everything; deleted means deleted.

Commentary has attached dramatic revenue numbers to the shutdown. We could not trace them to any primary source, so we will not repeat them. The structural read needs no leaked financials: a consumer video app lost to infrastructure. The distribution that won August is APIs, routers, ComfyUI graphs, and downloadable weights, and the labs still standing in video are the ones selling picks to builders rather than feeds to consumers.

The desktop is the new datacenter

Two papers this window put frontier-adjacent video capability on hardware creators already own. ABot-World-0 (arXiv 2607.19191) is an action-conditioned world model you steer with a keyboard: 720p at up to 16 fps on a single RTX 5090, 1.2 seconds from action to first frame, about 19 GiB of VRAM. The trick is distilling a bidirectional teacher that sees whole clips into a causal student that renders frame by frame, plus a “LongForcing” scheme that keeps long rollouts from drifting. Genie-3-class interactivity required a DeepMind datacenter and a subscription; this runs beside your ComfyUI graph.

JoyAI-Video-Edit (arXiv 2608.03974) is the mechanism behind where editing UX is going: a 16B autoregressive diffusion model that edits video as a stream, chunk by chunk, no known clip length required, at roughly 30 fps at 720p on a single B200. Its source-anchored distillation keeps edits faithful to the input footage over long horizons. Every “talk to your footage” product demo of the past month is some version of this loop. Between H3 quantized for local inference, LTX-2.5’s RTX tuning, and a playable world model on a 5090, the pattern holds: capability that needed a cluster in March fits under a desk in August.

If you build

• Treat September 24 as a hard incident date. Sora 2 API calls stop and data deletes. Migrate the integration and archive outputs this sprint, not next quarter.
• The abstraction layer stopped being advice and became a product category. We told you in July to put a layer between your product and any one video API. Runway’s Media Router is that layer off the shelf; the 3-arenas-3-leaders data is why it exists. Build or buy one either way.
• Add license review to the eval checklist. H3’s jurisdiction clause and LTX’s ARR cap are shippable for some teams and fatal for others, and the difference is a paragraph of legal text, not a benchmark score.
• Price local inference against your API bill. A 33B open model with day-one ComfyUI nodes turns batch-heavy workloads (variations, dailies, previz) into a GPU amortization problem instead of a per-clip line item.

If you create

• Pick models per task, per week. Text-to-video, image-to-video, and editing have different leaders separated by noise. Keep a current note of which arena leads where, and stop identifying as a one-model shop.
• Reference-driven consistency is the converged interface. H3’s Ref2VA takes 9 image references plus video and audio; xAI shipped 7-reference control with voice matching the same week. The consistency workflow transfers across all of them, and it is the skill that outlives any single model.
• Multishot generation moves the edit upstream. LTX-2.5 emits a multi-cut sequence with held characters from one prompt. Writing shot structure into the prompt is storyboarding; creators who think in sequences get more out of it than creators who think in clips.
• Previz on your own GPU is now real. An interactive world model on a single consumer card means you can scout blocking and camera paths at your desk, before spending API budget on finals.

Keeping score

In July we made a falsifiable claim: FLUX 3 Dev would ship materially behind the API version, and open weights would trail frontier generation by months. The first half still stands unresolved; BFL has shipped nothing open. The second half took real damage. H3 is 5 Elo off the closed text-to-video leader with weights anyone can pull, and it leads an arena outright. The squeeze we described was real in July and the catch-up in August was just as real; both can be true in a field moving this fast. What would change our mind back: the arenas drifting apart again as closed labs ship their next generation, or H3’s license friction turning out to gate the users who matter. We will keep reading the fine print either way.