The Best Open-Weight Video Models for Every Task in 2026

Key takeaways

• LTX-2.5 is the best open-weight video model most ad teams can use. It makes video and synchronized audio from text or an image, renders up to 4K, and leads the open models a US or EU team is licensed to run. Companies with $10 million or more in annual revenue need a paid license from Lightricks.

• The highest-rated open model is off-limits in the US and EU. MiniMax H3 tops the open-weight models on the Artificial Analysis leaderboards, but its license excludes the US, the EU, the UK, and South Korea.

• Wan 2.2 is still the safest all-rounder. It's Apache 2.0, so there are no revenue thresholds or territory limits, and its 5B version runs on a 24 GB graphics card.

• Pick by task, not by leaderboard. Talking heads, product consistency, editing, and long clips each have a specialist model that beats the general ones at that job.

• Read the license before the README. The licenses below range from MIT and Apache 2.0 to revenue thresholds and country exclusions, and that's often what decides which model a brand can put in a paid ad.

Why ad teams look at open-weight video

Open-weight models publish their trained weights, so you can download and run them on your own hardware or a rented GPU instead of paying per clip through someone else's service. For an ad team, that buys three things: no per-generation fees once the hardware is paid for, footage and product images that never leave your own machines, and the option to fine-tune a model on your brand's products or style.

The cost is work. Most of these models need a data-center GPU to run at full quality, someone has to install and maintain them, and none of them come with the reference tools, templates, or approvals of a finished product. The best commercial video models also stay ahead on quality: on the Artificial Analysis text-to-video leaderboard, the top closed models score above 1,150, and the best open model a US team can license, LTX-2.5, scores 947.

Several big 2026 releases aren't open at all. Alibaba has published no weights for anything after Wan 2.2, so its newer Wan versions are API-only, and many of the leading commercial video models don't publish weights.

How we picked

We matched each model to the job it does best for ad and ecommerce work: animating product shots, talking heads and lip sync, keeping a product or character consistent, editing existing footage, long clips, quick drafts on a single GPU, and camera control. For each, we read the maker's own model card, code repository, and license, and we used the Artificial Analysis Video Arena where it ranks the model. Quality claims that come only from a maker's own tests are labeled as such. The license weighed heavily: a model a brand can't use commercially in its market doesn't get a pick.

The picks at a glance

1. LTX-2.5 - Best for text-to-video, image-to-video, and audio

LTX-2.5 is Lightricks' 22-billion-parameter video model, released in August 2026.

What works: It generates video with synchronized audio from text or a still, so a product shot can become a clip with sound in one pass. Output runs at 24 fps by default, with temporal upscaling to 48 or 96 fps and a 4K setting of 3840 x 2176. The pipelines go past plain generation: audio-to-video, Retake to regenerate one stretch of a clip, Dub-It for lip-matched re-dubbing, and HDR output. On the Artificial Analysis leaderboards it scores 947 for text-to-video and 916 for image-to-video, the best of the open models a US or EU team can license. The previous version, LTX-2.3, still hosts control LoRAs for following drawn motion tracks and depth-style guides.

What doesn't: Companies with annual revenue of $10 million or more need a paid license for any production use, which covers most brands and many agencies. The download is about 66 GB, Lightricks gives no minimum VRAM figure, and the license bars using it commercially to train other models.

Best for: Small brands and studios under the revenue line, or larger teams willing to buy the commercial license, that want the strongest open model with sound.

2. Wan 2.2 - Best for a fully permissive all-rounder

Wan 2.2 is Alibaba's open video family, released in July 2025 and still its newest open base model.

What works: Apache 2.0 means no revenue line and no territory list, and the model card says Alibaba claims no rights over what you generate. There are text-to-video and image-to-video models at 480p and 720p, plus a smaller 5B model that does both at 720p and 24 fps and runs on a 24 GB card such as an RTX 4090. The family has grown since launch: a speech-to-video model, Wan-Animate-2 for moving a character with a driving video, and a music-driven dance model. Because so many other open models build on Wan, its tooling and fine-tunes are the widest of any open video model.

What doesn't: The full-quality 14B models need a GPU with at least 80 GB of VRAM, and Alibaba's own figure for the 5B model is under nine minutes per five-second 720p clip on a consumer card. There's no native audio in the main models and no current leaderboard score.

Best for: Brands and agencies that need a model their legal team will sign off on without a second read.

3. Kandinsky 6.0 - Best for drafts on a single consumer GPU

Kandinsky 6.0 is Kandinsky Lab's video model, open-sourced on October 6, 2026, in a 3B Lite version and a 29B Pro version.

What works: Both versions generate video with synchronized 44 kHz audio, including lip sync, under the MIT license. Kandinsky Lab's own timing table puts the distilled Lite model at 68 seconds for a five-second clip with audio on an RTX 4090, and about three minutes on a 16 GB RTX 5060 Ti. That's fast enough to rough out storyboards and hook ideas on a desktop card. First and last frame, audio-driven, and video-driven modes were added on October 25, and a super-resolution step takes clips to 1920 x 1080.

What doesn't: Every clip is five seconds. The base resolution is 480 x 864 before upscaling, and there's no third-party leaderboard score yet, so the quality claims are Kandinsky Lab's own.

Best for: Teams that want quick drafts with sound on the hardware they already have, under the simplest license on this list.

4. LongCat-Video-Avatar-1.5 - Best for talking heads and lip sync

LongCat-Video-Avatar-1.5 is Meituan's audio-driven avatar model, released on May 21, 2026, from the team behind the earlier MultiTalk and InfiniteTalk models.

What works: Give it a face and a voice track and it animates the speaker, with support for two speakers in one clip. It outputs at 480p or 720p, uses an eight-step distilled model for speed, and has an INT8 option to cut memory use. Meituan lists ecommerce marketing and commercial promotion among its intended uses, and it's MIT-licensed. For re-dubbing a video you already have, the same team's InfiniteTalk (Apache 2.0) syncs new audio onto existing footage and supports long durations.

What doesn't: Meituan gives no VRAM figure, and its quality results come from its own human evaluation, not a public leaderboard. A synthetic spokesperson in an ad can also trigger disclosure rules, such as New York's synthetic performer law.

Best for: UGC-style explainers, spokesperson clips, and localized versions of a talking-head ad.

5. SkyReels-V3 - Best for keeping a product or character consistent

SkyReels-V3 is Skywork's set of 14B video models, released in January 2026 for reference-to-video, video extension, and talking avatars.

What works: The reference-to-video model takes one to four reference images (a character, an object, a background) and builds a clip around them, and Skywork names product demonstrations and live-stream commerce as uses. That's the job ad teams most often need from an open model: the same bottle, shoe, or presenter in every shot. The extension model continues a clip by 5 to 30 seconds, and a talking-avatar model follows audio of up to 200 seconds. A low-VRAM mode with FP8 weights and offloading runs on cards with under 24 GB.

What doesn't: Reference-to-video works best at five seconds and 720p. Commercial use is allowed, but under the separate Skywork Community License, so read its terms before a campaign.

Best for: Product demos and recurring characters where the item has to look the same from clip to clip.

6. Wan2.1-VACE - Best for editing and inpainting existing footage

Wan2.1-VACE is Alibaba's all-in-one editing model, released in May 2025 under Apache 2.0.

What works: Alibaba describes it as an all-in-one model for video creation and editing, so one set of weights handles generating from references and reworking footage you already shot, rather than a separate model per task. It's still Alibaba's only open editing model. The 14B version runs at 480p or 720p, a 1.3B version runs at 480p, and the Apache license puts no limits on commercial use.

What doesn't: It's built on Wan 2.1, a generation behind, and Alibaba hasn't released a Wan 2.2 version. Alibaba gives no VRAM figure for the 14B editing model.

Best for: Fixing and repurposing existing product and campaign footage without a reshoot.

7. LongCat-Video - Best for long clips

LongCat-Video is Meituan's 13.6B base video model, released in October 2025 under MIT.

What works: It handles text-to-video, image-to-video, and continuation, and Meituan trained it natively on continuing videos. The maker says that lets it produce minutes-long clips at 720p and 30 fps without color drift or quality loss. For a product walk-through or a looping in-store screen, that's longer than any other open model here produces in one run.

What doesn't: The minutes-long claim is Meituan's, without a third-party benchmark, and there's no VRAM figure from the maker.

Best for: Longer product explainers, in-store loops, and background video where length matters more than peak detail.

8. Hy-Nomas - Best for camera control

Hy-Nomas is Tencent's video world model, released on October 21, 2026.

What works: It takes an explicit camera path along with text instructions and keeps generating as the camera moves, at 720 x 1280 and 24 fps. That makes it useful for planning a camera move around a set or a product display before a shoot. A documented single-RTX 4090 setup uses INT8 weights and CPU offloading, and the license is Apache 2.0 for the model, with third-party components listed separately.

What doesn't: It's built for exploring a generated scene, not for polished product shots, and the standard setup uses eight GPUs. The browser demo also relies on a separate prompt-enhancement service.

Best for: Previsualizing camera moves and spatial concepts.

The best open model you probably can't use

MiniMax H3, released in July 2026, is the strongest open-weight video model by a wide margin on the Artificial Analysis leaderboards: 1,161 for text-to-video and 1,128 for image-to-video. It makes 4 to 15 second clips at 768p with stereo audio, and its reference mode takes up to nine images, three videos, and three audio clips.

Its license, though, excludes the United States, the European Union, the United Kingdom, and South Korea, and businesses above $20 million in yearly revenue need MiniMax's authorization. For most readers of this site, it's out of bounds for client work.

Tencent's HunyuanVideo-1.5 has a similar catch. It's a capable 8.3B model with a 14 GB minimum when offloading, but its community license doesn't apply in the EU, the UK, or South Korea.

Which one fits your team

• You want the best quality you can license in the US or EU: LTX-2.5, with a paid license if you're over $10 million in revenue.

• Legal wants the simplest possible license: Wan 2.2 (Apache 2.0) or Kandinsky 6.0 (MIT).

• You only have a gaming GPU: Kandinsky 6.0 Lite, or Wan 2.2's 5B model on a 24 GB card.

• You need a spokesperson or a dubbed version: LongCat-Video-Avatar-1.5, or InfiniteTalk for existing footage.

• The product has to look identical in every shot: SkyReels-V3 reference-to-video.

• You're fixing footage you already have: Wan2.1-VACE.

Frequently asked questions

What is the best open-source video model in 2026?

On the Artificial Analysis leaderboards, MiniMax H3 is the top open-weight model, but its license excludes the US, the EU, the UK, and South Korea. LTX-2.5 is the best open model teams in those markets can license, and Wan 2.2 is the strongest fully permissive option under Apache 2.0.

Can I use open-weight video models in commercial ads?

Often, but check each license. Wan 2.2, Wan2.1-VACE, and InfiniteTalk use Apache 2.0, and Kandinsky 6.0 and the LongCat models use MIT. LTX-2.5 needs a paid license for companies with $10 million or more in annual revenue, and MiniMax H3 and HunyuanVideo-1.5 exclude whole regions.

What GPU do I need to run an open video model?

It depends on the model. Kandinsky 6.0 Lite runs on a 16 GB or 24 GB consumer card, and Wan 2.2's 5B model needs 24 GB. Wan 2.2's full 14B models need 80 GB, and several 2026 models are built for multi-GPU servers.

Which open video model generates audio?

LTX-2.5, Kandinsky 6.0, and MiniMax H3 generate synchronized audio with the video. LongCat-Video-Avatar-1.5 and InfiniteTalk work the other way, animating a face to match an audio track you supply.

Is Wan 3.0 open source?

No. Alibaba hasn't published weights for anything after Wan 2.2, so newer Wan versions are available only through its API.

Are open-weight video models as good as commercial ones?

Not yet at the top end. On the Artificial Analysis text-to-video leaderboard, the best closed models score above 1,150, against 947 for LTX-2.5. Open models win on cost at volume, privacy, and fine-tuning on your own products.