Best Open-Source AI Video Models, Measured (2026)

The open-source AI video model the community is building on fastest is one that creators in the United States, the European Union, the United Kingdom and South Korea are not licensed to run. MiniMax H3 drew 150 of the 311 community LoRAs, quantizations, finetunes and merges published for open video models since 31 August, 48% of the total, and its license names all four as Excluded Territories.

We pulled data from four public sources: the Hugging Face Hub API for 13 open video model families (the official repos, the Comfy-Org repackages, and all 2,433 community models that declare one of them as their base in the model tree), the license file of every model, every model card, and the ComfyUI source code. Here is what the numbers say about which open video model is actually worth your GPU, and which ones you can legally use.

Key Findings

1. MiniMax H3 took half of all new community work in two months

MiniMax published the H3 weights on 3 August. By the end of September, 390 community models were built on it, and 150 of them appeared in September alone. That is more new community models than Wan 2.1, Wan 2.2 and LTX-2.5 added in the same month combined (145). H3 also took 75 of the 119 new LoRAs, 63% of all LoRA training on open video models.

The downloads point the same way. The Comfy-Org repackage of H3 logged 22.1 million downloads in the month and MiniMax's own repo 3.6 million, together 62% of the monthly downloads across all 13 families. The biggest community builds are speed and memory work: Minimax-h3-Turbo (1.54 million downloads), the Unsloth GGUF quantizations (1.34 million) and FastH3, the 4-step distillation we covered in August (906,000).

The pull is easy to explain. H3 generates video with native stereo audio at up to 2K and up to 15 seconds, per its model card, ComfyUI supports it natively, and Comfy-Org publishes a repackaged copy ready to drop into a workflow. We covered the release in MiniMax H3 brings 2K AI video with native sound.

2. Its license excludes the US, the EU, the UK and South Korea

Section I of the MiniMax H3 Community License Agreement reads: "Excluded Territories means the European Union, the United Kingdom, the Republic of Korea and the United States of America." Section V adds that you may not use, reproduce, modify, distribute or display the H3 weights "or any of their Outputs or results" outside the Applicable Territory. The license defines "MiniMax H3 Works" to include every model derivative, so the 390 community builds carry the same terms.

MiniMax explains the carve-out in a license Q&A in the repo: video regulation in those four regions is still moving, and MiniMax is party to copyright litigation over generative video in the US. It calls the limit "not yet", not "not ever". Organizations in the excluded regions can apply for a separate license, and the hosted H3 API stays available worldwide. The license has two more terms that matter to anyone shipping a product: companies above $20 million in yearly revenue need written authorization, and any commercial product that uses H3 must display "MiniMax H3" in its interface.

H3 is not alone. The Tencent Hunyuan Community License on HunyuanVideo 1.5 excludes the EU, the UK and South Korea, though not the US, and both licenses bar you from using outputs to improve any other AI model. Together, the two region-limited families account for 151 of the 311 new community models in September, 48.6%. A creator in Los Angeles or London who follows the momentum on Hugging Face is following it onto weights they have no license to run.

3. Wan is still the foundation: two thirds of everything built

Momentum belongs to H3, but the installed base belongs to Alibaba. Wan 2.2 (July 2025) has 936 community models and Wan 2.1 (February 2025) has 702. That is 1,638 of the 2,433 community models across all 13 families, 67%. For LoRAs the lead is wider: 816 of the 1,077 community LoRAs, 76%, target a Wan model. Wan 2.2 alone has 459, more than three times H3's 139.

The older model is not fading either. Wan 2.1, now 19 months old, gained 63 new community models in September, more than LTX-2.5 and LTX-2.3 together (37). The most used community builds are packaging for local runs: Kijai's WanVideo_comfy logged 1.36 million downloads in the month and the QuantStack Wan 2.2 GGUF 465,000.

Wan also has the cleanest license among the popular families: Apache 2.0, no territory list, no revenue cap. It is likely the last open Wan for a while. Alibaba's newer generation ships as hosted API nodes, not weights, as we found when Wan 3.0 arrived in ComfyUI.

4. No ComfyUI support, no community

Six of the 13 families have no model-specific support in ComfyUI core, checked against supported_models.py on the master branch on 30 September: LongCat-Video, MOVA, MAGI-2 preview, SkyReels V3, SANA-Video 2.0 and Cosmos3 Image2Video. The best of them, LongCat-Video, has 7 community models. Every family with more than 20 has core support.

Support is necessary, not sufficient. Kandinsky 5.0 Video is in ComfyUI core under an MIT license and has one community model. But the pattern is strong enough to use as a filter: when a new open video model ships, whether ComfyUI supports it natively predicts whether anyone will build LoRAs for it. For H3, the Comfy-Org repackage logged six times the downloads of MiniMax's own repo, which says where people actually run these models (Hugging Face counts downloads differently depending on how a repo is packaged, so read that ratio as direction, not precision).

5. Size is the other wall: MAGI-2 preview has zero community models

Sand AI's MAGI-2 preview has everything on paper: Apache 2.0, synchronized audio and video, 10-second clips up to 1080p. It is also a 114-billion-parameter model (6 billion active per token) whose model card lists a 307 GB download and eight NVIDIA Hopper GPUs. Eight weeks after release it has zero community models. NVIDIA's Cosmos3 Image2Video, at 64.6 billion parameters by the Hub's count and aimed at robotics and physical AI, also has zero.

The families that won are the ones a creator can run on one card. Wan 2.2 ships a 5-billion-parameter TI2V model that its card says runs on consumer GPUs such as the RTX 4090, and HunyuanVideo 1.5 is 8.3 billion parameters with a stated minimum of 14 GB of GPU memory with offloading. Size alone is not the whole story: SANA-Video 2.0 is also 5 billion parameters and has zero community models, but it has no ComfyUI core support either.

6. LTX-2.5 has not pulled the community off LTX-2.3 yet

Lightricks shipped LTX-2.5 in August with a diffusion-based decoder, native multishot generation and synchronized audio, which we walked through in LTX-2.5: open-weights video and audio, now in ComfyUI. The community has been slower to move. LTX-2.3 still has 291 community models against 82 for LTX-2.5, and still logs more monthly downloads: 1.85 million against 1.67 million.

The switch has started. In September LTX-2.5 gained 22 new community models against 15 for LTX-2.3, and Lightricks has published 11 of its own LoRAs and IC-LoRAs for 2.5. One difference in the path to the weights: the LTX-2.5 repo is gated behind a Hugging Face login and license acceptance, and LTX-2.3 is not. Both sit under the same LTX-2 Community License, free for companies under $10 million in annual revenue, affiliates included.

Trend Analysis

Rising: MiniMax H3, with 150 new community models and 75 new LoRAs in one month. Audio-native video is where the attention goes: the two families with native audio that also have ComfyUI core support, H3 and LTX, took 187 of the 311 new community models.

Stable: Wan 2.1 and Wan 2.2 as the LoRA base. Together they gained 123 new community models in September and hold 76% of all community LoRAs.

Emerging: Speed distillations. Three of H3's five most downloaded community builds (Minimax-h3-Turbo, FastH3 and the Unsloth quantizations) exist to make a 33-billion-parameter model fit on smaller hardware or run in fewer steps.

Stalling: research-lab releases without ComfyUI support. MOVA, MAGI-2 preview, SkyReels V3, SANA-Video 2.0 and Cosmos3 Image2Video have one community model between them. HunyuanVideo 1.5 has core support but gained one new community model in September.

Predictions

1. MiniMax H3 passes Wan 2.1 in total community models by the end of January 2027. H3 has 390 and added 150 in September; Wan 2.1 has 702 and added 63. At those rates H3 reaches about 990 by the end of January against about 954 for Wan 2.1.

2. LTX-2.5 passes LTX-2.3 in monthly downloads by the end of October. The gap is 1.85 million to 1.67 million, and new community work already favors 2.5 by 22 to 15.

3. No open video model without ComfyUI core support reaches 20 community models in 2026. The best today is LongCat-Video with 7, eleven months after release.

4. MAGI-2 stays under 10 community models until Sand AI ships a checkpoint that runs on one GPU. A permissive license has not been enough at 307 GB and eight Hopper cards.

What This Means for Creators

If you work in the US, the EU, the UK or South Korea: do not build a client pipeline on the H3 weights or on any of the 390 models built from them. Use H3 through MiniMax's hosted API, apply for a license, or pick a model whose license covers you. For audio-native video locally, that is LTX-2.5 (free under $10 million in revenue). For a permissive license with no conditions, it is Wan 2.2.

If you train or sell LoRAs: Wan 2.2 has the largest audience, 459 community LoRAs under Apache 2.0, and Wan 2.1 is still growing. H3 is where new LoRA work is happening fastest, but your buyers in the four excluded regions cannot use it.

If you have a single 12 to 16 GB GPU: start with Wan 2.2 TI2V-5B or HunyuanVideo 1.5 (14 GB with offloading, and not licensed in the EU, UK or South Korea). For H3 or LTX-2.5, use the community GGUF and distilled builds rather than the full weights.

If you build a product: read the commercial terms before the benchmarks. H3 requires "MiniMax H3" in your interface and authorization above $20 million in revenue; LTX needs a paid license from $10 million; both H3 and HunyuanVideo 1.5 forbid using outputs to train other models. The same questions came up for open lip-sync models in our license audit of open lip-sync tools.

If you are evaluating a new open video release: check two things before you download: whether ComfyUI core supports it, and how many models sit in its Hugging Face model tree. Six of 13 families never got past one.

Full Data

Every family with its license, ComfyUI status and adoption, measured on 30 September 2026. Community models are Hugging Face repos whose model card declares one of the family's official repos or its Comfy-Org repackage as base model, counted one level deep and deduplicated, excluding repos published by the model maker and by Comfy-Org. "New" means created from 31 August to 30 September.

Monthly downloads are the Hub's own figure for the official repos plus the Comfy-Org repackage. Hugging Face counts a download each time a repo's query file is requested (config.json by default, model_index.json or top-level safetensors for diffusers repos, every file for GGUF), so repos packaged differently are counted differently, and MAGI-2 preview's layout registers no count at all. Treat downloads as a rough signal. The community model counts are the cleaner measure of what people build on.

Frequently asked questions

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

It depends on what you need. By community adoption it is MiniMax H3, with 150 new community models in September, but its license excludes the US, the EU, the UK and South Korea. For a permissive license with the largest LoRA library it is Wan 2.2 (Apache 2.0, 459 community LoRAs). For native audio under a license that covers the US and EU, it is LTX-2.5, free under $10 million in annual revenue.

Can I use MiniMax H3 in the United States?

Not under the MiniMax H3 Community License. It lists the United States, the European Union, the United Kingdom and South Korea as Excluded Territories and forbids using the weights or their outputs there. MiniMax says organizations in those regions can apply for a separate license, and its hosted H3 API remains available worldwide.

Which open video model has the most LoRAs?

Wan 2.2, with 459 community LoRAs on Hugging Face, followed by Wan 2.1 with 357 and MiniMax H3 with 139. H3 is growing fastest: 75 new LoRAs in September against 18 for Wan 2.2.

Is LTX-2.5 free for commercial use?

Yes, for companies with less than $10 million in annual revenue, counting subsidiaries and affiliates. Above that, the LTX-2 Community License requires a paid commercial license from Lightricks.

Which open video models have fully permissive licenses?

Wan 2.2, Wan 2.1, MOVA, MAGI-2 preview and SANA-Video 2.0 use Apache 2.0. LongCat-Video and Kandinsky 5.0 Video use MIT, and NVIDIA's Cosmos3 models use OpenMDW 1.1. None of them has a territory limit or revenue cap.

Why can't I use HunyuanVideo 1.5 in the EU?

The Tencent Hunyuan Community License defines its territory as worldwide excluding the European Union, the United Kingdom and South Korea, and says any use outside that territory is unlicensed. Companies with more than 100 million monthly active users also need a separate license from Tencent.

This research was produced by Creative AI News.

Subscribe for free to get the weekly digest every Tuesday.