The Next Open-Weight Video Models: What to Expect in 2026 and Beyond

What open-weight AI video models exist now, what makers have promised next, how far they trail closed models, and the signals that show a release is close, from MiniMax H3 and LTX-2.5 to FLUX 3 Dev.

15 min read

15 min read

Close-up of a blue circuit board with chips and components

Key takeaways

  • The best open-weight video model is close to the best closed one. MiniMax H3 scored 1161 on Artificial Analysis's text-to-video arena, 14 points behind the closed Wan 3.0 at 1175.

  • The gap is one model deep. The next open model on the board, Lightricks' LTX-2.5, sits more than 200 points lower.

  • The only confirmed next release from a major lab is Black Forest Labs' FLUX 3 Dev, announced on July 23, 2026 with no date. It's still not out.

  • Alibaba's flagship Wan line has stopped opening. The newest open Wan weights are Wan 2.2 and the specialist Wan-Animate-2, and nothing official promises Wan 3.0 weights.

  • Most top open models carry revenue caps or territory limits, so check the license before building on one.

Where open-weight video stands today

Open weights here means you can download the model and run it yourself, which isn't the same as open source. Most of the strongest models come with community licenses that limit who can use them.

The main open-weight video models as of November 11, 2026.

Model

Released

Size

License

MiniMax H3

August 3, 2026

33B

MiniMax community license, with territory limits and a revenue threshold

Pika H3 Flex

October 29, 2026

33B

Same MiniMax license

Lightricks LTX-2.5

August 2026

About 22B

LTX-2.x community license, free under $10M annual revenue

Kandinsky 6.0 Video

October 2026

29B Pro, 3B Lite

MIT

Wan-Animate-2

August 2026

14B

Apache 2.0

Wan 2.2

2025

14B and 5B

Apache 2.0

HunyuanVideo-1.5

November 2025

8.3B

Tencent Hunyuan Community License

MiniMax H3 is the standout. MiniMax announced it on July 30 and posted the weights on Hugging Face on August 3. It makes 4 to 15 seconds of 768p video with stereo audio, and it runs in ComfyUI. On Artificial Analysis's open-weights board it leads, followed by Pika's H3 Flex, which is built on H3, then LTX-2.5.

How big is the gap with closed models?

Smaller than it looks, and thinner too. On Artificial Analysis's text-to-video arena, the closed Wan 3.0 ranks first at 1175, and the open MiniMax H3 ranks third at 1161. Pika H3 Flex scores 1124. After those two, which share a base model, the best open entry on the board is LTX-2.5 Fast at about 943. In practice, one maker's model is carrying the open side, so a single change of plans there would reshape the picture.

What the makers have said about what's next

Only a few statements are on the record, and every one lacks a date.

  • Black Forest Labs. When it announced the FLUX 3 family on July 23, BFL said the planned capabilities, including an open-weight multimodal backbone called FLUX 3 Dev, would roll out "over the next few weeks and months." FLUX 3 Dev still hadn't shipped 111 days later. BFL did release the open FLUX 3 Image Klein on October 15, but that's an image model, and FLUX 3 Video and FLUX 3 Max stay closed.

  • MiniMax. The H3 model card says its sparse-attention code "will be released in a future update," and a 2K regeneration module isn't open yet.

  • Tencent. Most of HunyuanVideo-1.5's distilled and sparse variants have shipped, but the 720p text-to-video distill is still marked "coming soon," and no successor model has appeared.

  • Alibaba. The Wan-Animate-2 card mentions a Lite version for real-time use, but no weights are listed.

  • Lightricks. It hasn't published a roadmap, so any talk of an LTX-3 is guesswork.

Why Wan's flagship may stay closed

Wan 2.1 and 2.2 were released as Apache 2.0 open weights, and later numbered versions, 2.5 through 3.0, were released as hosted models only. Wan 3.0 now tops Artificial Analysis's board as a proprietary model. Some blogs claim Wan 3.0 weights exist, but Alibaba hasn't released them and none appear in the Wan-AI Hugging Face organization. Alibaba's recent open releases are specialist models like Wan-Animate-2 and Wan-Dancer, so the pattern points to a hosted flagship with open side models. That's an inference, not a statement from Alibaba.

Signals that a release is close

Most releases this year showed one or more of these signs first.

  • A date in the announcement. MiniMax said the weights would come "within days" and delivered in four.

  • A Hugging Face repo. Wan-Animate-2's files appeared on the Hub before most coverage did.

  • A ComfyUI workflow. H3 and LTX shipped with ComfyUI workflows, and Pika's card links them too, so a Comfy-Org repo is a strong sign.

  • A paper on arXiv, as with Kandinsky 6.0.

  • A "coming soon" line in a README is weak evidence. Tencent's has sat there for months.

An announcement without a date is the weakest signal. FLUX 3 Dev shows it can stay unreleased for months.

What to expect

These are predictions, not confirmed plans.

  • More models built on H3. Pika's Flex release shows MiniMax's weights are already a base for others.

  • Smaller and faster versions. Makers keep shipping distilled variants, such as Pika's six-step Flex, Kandinsky's 3B Lite, and LTX-2.5 Fast.

  • FLUX 3 Dev, eventually. BFL has committed to it, so it's the likeliest big open release, but it's unwise to plan around a date.

  • Tighter licensing. The top open models are the most restricted, while permissive Apache 2.0 or MIT licenses show up mostly on smaller or specialist models.

What hardware you'll need

Few makers give a hardware figure. HunyuanVideo-1.5 lists a 14GB minimum with offloading, and Wan 2.2's 5B model fits a 24GB card. For the big new models, MiniMax shows H3 running on 4 GPUs, Pika tested on an 80GB H100, and Wan-Animate-2 is tuned for 8 A800s. If you're on a single consumer card, the smaller and quantized variants are the realistic path for now.

How to track new releases

  • Watch the Hugging Face pages for Wan-AI, Lightricks, MiniMaxAI, black-forest-labs, kandinskylab, and Pika-Labs, and sort the text-to-video model list by newest.

  • Follow the GitHub repos for Wan, HunyuanVideo, and LTX, and ComfyUI's release notes.

  • Check Artificial Analysis's open-weights leaderboard, which adds new models quickly.

  • Read each maker's own blog, such as bfl.ai/blog, since model cards often carry the real license terms.

Frequently asked questions

Which open-weight video model is best right now?

MiniMax H3, on Artificial Analysis's text-to-video arena, where it scores 1161 and sits third overall. Check its license first, because it excludes the EU, UK, South Korea, and US and requires permission above $20M in annual revenue.

Is Wan 3.0 open source?

No. Alibaba hosts it, and it hasn't released weights or promised them. The newest open Wan weights are Wan 2.2 and Wan-Animate-2.

When will FLUX 3 Dev be released?

BFL hasn't said. It announced FLUX 3 Dev on July 23 with no date, and it was still unreleased on November 11.

Can I run open video models on my own computer?

Some, yes. HunyuanVideo-1.5 needs at least 14GB of video memory with offloading, and Wan 2.2's 5B model runs on about 24GB. The largest models need data-center cards.

Are open-weight models free for commercial use?

Not always. LTX-2.5 is free below $10M in annual revenue, and MiniMax H3 excludes the EU, UK, South Korea, and US from its licensed territory and requires written permission above $20M in annual revenue. Wan 2.2 and Wan-Animate-2 use Apache 2.0.

How do I know a release is coming?

Look for a Hugging Face repo, a ComfyUI workflow, or a dated announcement. A "coming soon" note or an undated promise isn't much evidence.

Explore Topics

Icon

0%

Explore Topics

Icon

0%