Open-Source Video Models Are 262 Elo Behind. One Weight Drop Closes It.

This is the fifth entry in our open-source AI video series. Part four ended with a line we expected to sit on for months: if Alibaba opens Wan 2.7, that becomes the story. It has not happened yet.

Nine days later, two other labs made the same promise. Black Forest Labs announced FLUX 3 on July 23 with an open-weight backbone to follow. MiniMax announced H3 on July 31 and said the weights ship "in the coming days." Neither has published a file.

That sounds like more of the same, and in one sense it is. But the arithmetic underneath changed, and it is worth looking at directly, because the distance between open and closed video is now small enough to be measured in a single number.

TL;DR

• The open-weights video gap is 262 Elo. The best downloadable text-to-video model with audio is LTX-2.3 Fast at 980 on the Artificial Analysis arena. MiniMax H3, announced last week, sits at 1242.
• Three labs are holding the trigger. FLUX 3 Dev, MiniMax H3, and Wan 3.0 are all announced as open-weight. As of today, none of the three exists on Hugging Face.
• One weight drop closes almost the entire gap. If MiniMax ships what it announced, open-weight video goes from "good enough to iterate with" to second place in the world overnight.
• The catch is the license, not the capability. Nobody at the frontier is shipping Apache 2.0 anymore. Revenue-capped community licenses are the new default.

The 262-Elo gap in open-weight video models, measured

Artificial Analysis runs a blind arena for video: two clips from the same prompt, users vote, Elo comes out the other end. It has a filter for open weights, which makes it the cleanest available read on how far behind downloadable models are.

Comparing like for like on the with-audio board, because H3 and FLUX 3 both generate native audio:

Two things stand out. The top ten is entirely closed or unshipped. And the entire open-weights leaderboard, every position of it, is Lightricks. There is no second open lab in the running right now.

1242 minus 980 is 262. That is the number this article is about. It is the amount of quality that would transfer from the closed column to the open column if exactly one company followed through on exactly one announcement.

For scale: 262 Elo in a preference arena is not a rounding error and not an unbridgeable chasm. It is roughly the distance between "usable for previz and social clips" and "usable for the shot you actually deliver."

Three promises, zero files

Part four published a one-command check for whether a model is genuinely released. Here it is pointed at all three claims at once, run on the date below:

for q in Wan3 Wan2.7 MiniMax-H3 FLUX.3; do
  echo -n "$q -> "
  curl -s "https://huggingface.co/api/models?search=$q&limit=5" \
    | python3 -c "import sys,json; d=json.load(sys.stdin); print(len(d), [m['modelId'] for m in d])"
done

Every one of them comes back empty, or with unrelated community repos that happen to match the string. Checking what the three labs have actually published, newest first:

That last row is the contrast worth sitting with. While three labs were announcing open weights, Lightricks pushed 17 repository updates in July alone, including eight new IC-LoRAs in the final four days of the month: dubbing, relighting, in/outpainting, spatial upscaling, deblur, decompression, water simulation, and one that shaves beards. Announcements are cheap. Adapters are not.

What each drop would actually change

MiniMax H3: the one that moves the number

H3 is a multimodal video model that takes text, images, video, and audio in one context and returns up to 15 seconds at 2K with native stereo sound. Artificial Analysis ranks it first in the world for video editing and second for text-to-video with audio.

MiniMax has committed to the MiniMax Community License: free for non-commercial use, free for commercial use by organizations under 20 million dollars in annual revenue, with attribution required. As of today there is no H3 repository on the MiniMaxAI org.

If it ships, this is the whole story. Everything else in this article is a footnote to it.

FLUX 3: the one that changes the shape

FLUX 3 is Black Forest Labs' first video model, and it is not only a video model. It is jointly trained on image, video, audio, and action prediction in one architecture, which is why it also powers FLUX-mimic for robotics. Video output runs up to 20 seconds with native audio.

The open-weight piece is called FLUX 3 Dev, described in BFL's own announcement as "open-weight access to a multimodal backbone, for content creation (video, audio and image) and action prediction." What the announcement does not contain is a date or a license name.

Worth being precise about what this would mean if it lands. It would be the first downloadable model that generates video, audio, images, and robot actions from one set of weights. That is a different kind of release from H3, and for anyone building agents that need to both see and act, it is arguably the more interesting one. It is also the vaguest of the three promises.

Wan 3.0: the one to discount

Wan 3.0 is pre-announced as an Apache 2.0 open-weight release for mid-2026. Treat that with more caution than the other two, for reasons that are entirely evidence-based.

We could not find a primary Alibaba source for Wan 3.0 at all. The specifications circulating contradict each other, with the same "confirmed" model described as 60B dense in one place and a 27B mixture-of-experts with 14B active in another. Those pages trace back to single-purpose domains built around the model name, the same pattern part four documented for Wan 2.7.

The track record does not help either. Wan 2.5 was pre-announced as open and never appeared. Wan 2.6 shipped closed. Wan 2.7 is on the arena at 1163 and remains API only. Official Wan open weights still stop at 2.2, where they have been since August 2025.

None of that makes Wan 3.0 fake. It makes it unverifiable, which for planning purposes is the same thing.

The licenses are the real story

The reflex reading of an open-weights wave is that the field is getting more permissive. It is getting more available, which is not the same thing.

The only genuinely permissive entry is the oldest one. Every model that has landed or been announced since is a revenue-capped community license, and the caps come with details worth reading before you assume you are under them. The LTX threshold counts subsidiaries, affiliates, and companies under common control collectively, so a small studio inside a large group may already be over it.

For most people reading this, a 10 or 20 million dollar cap is functionally no cap. That is exactly why it works as a strategy, and exactly why it is worth naming: the industry has found a license that feels open to almost everyone who downloads it while staying commercially closed to everyone who matters to revenue.

The lawsuit nobody is pricing in

One complication sits underneath the MiniMax promise specifically.

Disney, Marvel, Lucasfilm, Twentieth Century Fox, Universal, DreamWorks Animation, and Warner Bros. Discovery sued MiniMax on 2025-09-16 in the Central District of California, alleging Hailuo was trained on unauthorized copies of their work and reproduces recognizable characters on simple prompts. On 2026-05-26, Judge Stanley Blumenfeld denied MiniMax's motion to dismiss, finding the studios had plausibly alleged widespread intentional infringement. The case is live.

So the scenario on the table is a lab open-sourcing the weights of a video model whose predecessor is under active infringement litigation from most of Hollywood. We are not aware of a precedent for that, and the honest answer about downstream exposure for people who fine-tune and ship products on those weights is that nobody knows yet.

Factor it into planning, not into panic. It is a reason to read the license text when it lands rather than a reason to skip the download.

What you can actually run today

One model in this entire article is downloadable, and it is the same one as last time. LTX-2.3: 22 billion parameters, synchronized audio and video from a single model, 2.13 million downloads on the base repository.

The shortest path from nothing to a clip:

pip install -U diffusers transformers accelerate

python3 - <<'PY'
import torch
from diffusers import DiffusionPipeline
from diffusers.utils import load_image, export_to_video

pipe = DiffusionPipeline.from_pretrained(
    "Lightricks/LTX-2.3", torch_dtype=torch.bfloat16, device_map="cuda"
)
image = load_image("your-still.png")
out = pipe(image=image, prompt="A man with short gray hair plays a red electric guitar.").frames[0]
export_to_video(out, "output.mp4")
PY

Budget 24GB of VRAM with the int8 build for 1080p work. Community reports put a 10-second 1080p clip at roughly 4 to 6 minutes on an RTX 4090. Two constraints will bite you on the first run: width and height must be divisible by 32, and frame count must be divisible by 8 plus 1. Ask for 100 frames and you get an error rather than a video.

If you are on Apple Silicon, our LTX-2 on Mac walkthrough still applies, and the hardware stack guide covers where a 22B model sits in memory terms.

What to watch in the next 90 days

Three checks, in descending order of how much they matter:

1. Does MiniMaxAI/MiniMax-H3 appear on Hugging Face? "Coming days" was July 31. If it is still empty at the end of August, the promise has entered Wan territory.
2. Does FLUX 3 Dev get a date and a license name? The capability claim is the most interesting of the three. The commitment behind it is currently the thinnest.
3. Does Lightricks stay alone at the top? If both promises slip, the open-weights leaderboard remains a single-vendor board for another quarter, and the 262 stays where it is.

The optimistic case here is real, and it is narrower than the coverage suggests. Open-weight video does not need three labs to follow through. It needs one. The gap has never been this close to closing on a single upload.

Set a reminder for two weeks out and run the check yourself. It takes ten seconds:

curl -s "https://huggingface.co/api/models?author=MiniMaxAI&sort=lastModified&direction=-1&limit=5" \
  | python3 -c "import sys,json; [print(m['modelId'], str(m['lastModified'])[:10]) for m in json.load(sys.stdin)]"

Sources and further reading

• Part four: what actually shipped in mid-2026, and what did not
• Part one: our tested open-source video model comparison
• Artificial Analysis open-weights text-to-video leaderboard
• Artificial Analysis text-to-video leaderboard (all models)
• Black Forest Labs: FLUX 3 announcement
• SCMP: MiniMax challenges ByteDance with open weights for H3
• Disney Enterprises, Inc. v. MiniMax, 2:25-cv-08768 (C.D. Cal.)
• Lightricks/LTX-2.3 on Hugging Face
• The LTX-2 Community License Agreement

Tested on: not independently tested. We did not run LTX-2.3, H3, or FLUX 3. Every release claim above was verified directly against the public Hugging Face API on the date below, using the commands published in this article: model existence, organization publishing history, download counts, and last-modified dates. Elo figures are read from the Artificial Analysis video arena on the same date and are user-vote preference scores, not capability benchmarks. Capability descriptions, license terms, and release intentions are the projects' own. VRAM and generation-time figures for LTX-2.3 are community-reported.
Date checked: 2026-08-01