# Best open-weight video generation models

> Source: https://aiwiki.ai/wiki/best_open_weight_video_models
> Updated: 2026-11-19
> Fact-checked: 2026-11-19
> Categories: AI Models, Generative AI, Open Source AI, Video Generation
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Best open-weight video generation models." aiwiki.ai, 19 Nov 2026. https://aiwiki.ai/wiki/best_open_weight_video_models
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Best open-weight video generation models** is a comparison of [AI video generation](https://aiwiki.ai/wiki/ai_video_generation) models whose trained weights can be downloaded and run on the user's own hardware, grouped by the licence that governs those weights. As of November 19, 2026, the highest-placed model flagged as open weights on the [Artificial Analysis](https://aiwiki.ai/wiki/artificial_analysis) text-to-video leaderboard was [MiniMax H3](https://aiwiki.ai/wiki/minimax_h3) (Elo 1159, third of 35 entries), whose licence excludes the European Union, the United Kingdom, South Korea and the United States.[1][42] Families released under the [Apache License 2.0](https://aiwiki.ai/wiki/apache_license), such as Alibaba's Wan 2.1 and Wan 2.2, carry no territorial or revenue conditions in their licence but do not appear on that leaderboard.[1][7][9] This article does not rank the models. It groups them by licence tier and by documented capability, and gives each leaderboard or download figure with its source and date.

The term [open weights](https://aiwiki.ai/wiki/open_weights) is used throughout. Some model cards use the phrase "open-source": Tencent's HunyuanVideo card calls the model an "open-source video foundation model" and the MOVA card describes a "fully open-source framework".[10][36] This article calls all of the models open-weight and describes each licence by its own terms. Still-image models are covered in [Best open-weight image models](https://aiwiki.ai/wiki/best_open_weight_image_models), and hosted video services in [Best AI video generators](https://aiwiki.ai/wiki/best_ai_video_generators).

## How the models were selected

The criteria below are this article's own. A model is listed when three conditions hold: its weights are published in a public repository (all of those listed are on [Hugging Face](https://aiwiki.ai/wiki/hugging_face)), the licence terms that apply to the weights could be read, and the model shows a public signal of evaluation or use, read on November 19, 2026. That signal is a position on a named leaderboard, a non-zero "Downloads last month" figure on the model's Hugging Face page or, for three repositories whose download counter showed 0 on that date, the number of likes on the page: sand-ai/MAGI-1 (611 likes), apple/FastFrames-L (210 likes, in a repository created three days earlier) and sand-ai/MAGI-2-preview (126 likes, and a leaderboard listing).[29][30][54] The list is not exhaustive, and releases whose licence terms were not confirmed on that date are left out.

The licence named in each row was read from a licence file in the weights repository where the repository has one. Several repositories have none. For LTX-2, LTX-2.3, LTX-2.5 and Cosmos 3 the model card metadata links to the licence text on GitHub or openmdw.ai, and that text was read.[18][19][38][70] For Wan 2.2, MAGI-1, MAGI-2 Preview, MOVA, Ovi, Open-Sora 2.0 and Step-Video-T2V the basis is the licence named in the model card or its metadata, together with the licence of the developer's code repository on GitHub.[9][28][29][30][32][35][36][37][73][74]

Hugging Face counts downloads on its servers from HTTP requests, including GET and HEAD requests, to a small set of query files such as `config.json`, a method intended to avoid counting one download of a multi-file model several times.[4] The figure counts requests to one repository, not users, and copies of the same weights in other repositories are counted separately. The Comfy-Org repackaging of Wan 2.2, for example, showed 9,837,533 downloads for the month, against 21,054 for Alibaba's own Wan2.2-TI2V-5B repository.[67][71]

Every repository listed could be downloaded without an access request except Lightricks/LTX-2.5, which is gated with automatic approval.[70] On Hugging Face, an access request shares the requester's username and email address with the model author unless the author has opted out of collecting it.[5] The public LTX-2.5 page says that visitors must agree to share their Hugging Face information and that accepting gives consent "to receive offers and updates including targeted and personalized advertisements".[70] The LTX-2.5 entry below is based on Lightricks's public GitHub repository, changelog and licence file rather than on the gated model card.[21][22]

Within each table the rows run from newest to oldest release. The order is chronological and is not a quality ranking. Release dates are the developer's stated date where one was found. For Pika H3 Flex, Hyperflow H3, Ovi, Mochi 1 preview and FastFrames-L, and for the LTX-2 and LTX-2.3 months, the date is the creation date of the weights repository on Hugging Face.[18][19][27][45][46][54][73] Parameter counts, resolutions and clip lengths are those stated by each developer, except that the LTX-2, LTX-2.3 and LTX-2.5 counts are read from the checkpoint file names (19b and 22b).[18][19][22]

## Models by licence tier

### Permissive licences

The models in this group are released under the Apache License 2.0, the [MIT License](https://aiwiki.ai/wiki/mit_license) or [OpenMDW](https://aiwiki.ai/wiki/openmdw) 1.1. ComfyUI's documentation for Wan 2.2 summarizes the Apache 2.0 terms as allowing use, modification and distribution, including for commercial purposes, as long as the copyright notice and licence text are retained.[59] OpenMDW 1.1 grants permission to deal in the model materials without restriction, requires that the agreement and notices be kept, and states that it places no restrictions on outputs.[40]

| Model | Developer | Weights released | Parameters | Modes | Documented output | Licence |
|---|---|---|---|---|---|---|
| [MAGI-2 Preview](https://aiwiki.ai/wiki/magi_2) | [Sand.ai](https://aiwiki.ai/wiki/sand_ai) | 2026-08-05 | 114B total, 6B active per token | Text or image to video with audio | 10 s clips, refiner up to 1080p | Apache 2.0 per the model card; see the note on a bundled component [30][31] |
| [NVIDIA Cosmos 3](https://aiwiki.ai/wiki/nvidia_cosmos_3) (Super, Nano, Edge) | [NVIDIA](https://aiwiki.ai/wiki/nvidia) | 2026-05-31; Edge 2026-07-20 | 64B, 16B, 4B | Super and Nano: text, image, video and action input; image, video, audio, action and text output. Edge: text, image and action input; image, video, action and text output | Super and Nano: 5 to 400 frames (189 by default). Edge: 50 to 150 frames at 256p or 480p, 12 to 30 fps | OpenMDW 1.1 [38][39][72] |
| MOVA | OpenMOSS | 2026-01-29 | 32B total, 18B active | Text, or image and text, to video with audio | 360p and 720p checkpoints | Apache 2.0 [35][36] |
| LongCat-Video | [Meituan](https://aiwiki.ai/wiki/meituan) | 2025-10-25 | 13.6B | Text-to-video, image-to-video, video continuation | 720p at 30 fps | MIT [33][34] |
| Ovi | Authors at [Character AI](https://aiwiki.ai/wiki/character_ai) and Yale University | 2025-09-30 | 11B | Text, or text and image, to video with audio | 5 s or 10 s at 24 fps, 960x960 | Apache 2.0 [37][73] |
| Wan 2.2 | [Alibaba](https://aiwiki.ai/wiki/alibaba) | 2025-07-28 | 27B (two-expert A14B models); 5B (TI2V-5B) | Text-to-video, image-to-video; later speech-to-video and character animation | 480P and 720P; TI2V-5B at 720P, 24 fps | Apache 2.0 [9] |
| [Wan 2.1 VACE](https://aiwiki.ai/wiki/wan_vace) | Alibaba | 2025-05-14 | 1.3B, 14B | Video creation and editing | 480P (1.3B); 480P and 720P (14B) | Apache 2.0 [8] |
| MAGI-1 | Sand.ai | 2025-04-21 | 24B, 4.5B | Text-to-video, image-to-video and video-to-video, generated autoregressively | 24-frame chunks generated in sequence | Apache 2.0 [29] |
| [Open-Sora](https://aiwiki.ai/wiki/open_sora) 2.0 | HPC-AI Tech (hpcaitech) | 2025-03-12 | 11B | Text-to-video, image-to-video | 256px and 768px | Apache 2.0 [28][74] |
| [Wan 2.1](https://aiwiki.ai/wiki/wan_2_1) | Alibaba | 2025-02-25 | 14B, 1.3B | Text-to-video, image-to-video, first-last-frame-to-video | 480P and 720P | Apache 2.0 [6][7] |
| [Step-Video-T2V](https://aiwiki.ai/wiki/step_video) | [StepFun](https://aiwiki.ai/wiki/stepfun) | 2025-02-17 | 30B | [Text-to-video](https://aiwiki.ai/wiki/text_to_video) | Up to 204 frames | MIT [32] |
| [Mochi 1](https://aiwiki.ai/wiki/mochi_1) preview | [Genmo](https://aiwiki.ai/wiki/genmo) | 2024-10-22 | 10B | Text-to-video | 480p | Apache 2.0 [27] |
| [CogVideoX](https://aiwiki.ai/wiki/cogvideo)-2B | [Zhipu AI](https://aiwiki.ai/wiki/zhipu_ai) (zai-org) | 2024-08-06 | 2B | Text-to-video | 720x480, 6 s, 8 fps | Apache 2.0 [23][25] |

Notes on this group:

- Wan 2.1: the GitHub repository and most model cards date the release of code and weights to February 25, 2025, while the Wan2.1-T2V-14B card says February 22, 2025.[6][7] The card states that Alibaba claims no rights over generated content and that users remain accountable for how they use the models.[6]
- Wan 2.2: the A14B text-to-video and image-to-video models split denoising between a high-noise and a low-noise expert, a [mixture-of-experts](https://aiwiki.ai/wiki/mixture_of_experts) design.[9] Wan2.2-S2V-14B (audio-driven video, August 26, 2025) and Wan2.2-Animate-14B (character animation and replacement, September 19, 2025) were added later under the same repository.[9] Alibaba's Wan-AI organization also published Wan-Dancer-14B (dance video generated from music, July 13, 2026) and Wan2.2-Animate-2-14B (August 7, 2026), whose cards state the Apache License 2.0.[76][77]
- MAGI-2 Preview: the model card and its metadata say Apache 2.0, and the weights repository contains no licence file.[30] The download includes a `stable-audio-open-1.0` audio VAE.[30] The upstream Stable Audio Open 1.0 repository is gated. Its metadata names the licence "stable-audio-community", and its public page lists the licence as the Stability AI Community License and refers commercial users to Stability AI's licence page.[66] The gated licence text was not read for this article, and the MAGI-2 card does not say which terms apply to the bundled file.
- Cosmos 3 is described by NVIDIA as a set of world models for physical AI (robotics, autonomous vehicles and smart spaces), not as a creative video tool, and its card says it is ready for commercial and non-commercial use.[38]
- CogVideoX-2B: the developers announced on August 27, 2024 that its licence had been changed to Apache 2.0; the larger CogVideoX models stayed under a separate licence (see below).[23]

### Conditional licences

These licences allow commercial use only under stated conditions such as revenue or user thresholds, territorial exclusions, registration, or limits on training other models.

| Model | Developer | Weights released | Parameters | Documented output | Licence | Main conditions on commercial use |
|---|---|---|---|---|---|---|
| [Pika H3 Flex](https://aiwiki.ai/wiki/pika_h3_flex) | Pika (Mellis, Inc.) | 2026-10-29 | Rank-128 [LoRA](https://aiwiki.ai/wiki/lora) adapters for MiniMax H3 | 480p and 768p, 5 to 15 s, stereo audio | MiniMax H3 Community License | As MiniMax H3 [45] |
| Hyperflow H3 | Video Rebirth | 2026-09-17 | 8-step LoRA (rank 256) for MiniMax H3 | As base model | MiniMax H3 Community License | As MiniMax H3 [46] |
| LTX-2.5 | Lightricks | 2026-08-11 | 22B (checkpoint file names) | Default 1024x1536 at 24 fps; 3840x2176 supported | LTX-2.x Community License (2026-08-11, amended 2026-10-29) | Entities with annual revenue of at least 10 million US dollars need a paid licence for any use outside the defined non-commercial purposes [21][22] |
| [MiniMax H3](https://aiwiki.ai/wiki/minimax_h3) | [MiniMax](https://aiwiki.ai/wiki/minimax) | 2026-08-03 | 33B (H3-Omni-Transformer) | 768p by default, 4 to 15 s, 24 fps, 32 kHz stereo audio | MiniMax H3 Community License (2026-08-02) | Not licensed in the EU, UK, South Korea or US; written authorization needed above 20 million US dollars in yearly revenue; "MiniMax H3" must be displayed in commercial products [41][42][44] |
| LTX-2 and LTX-2.3 | Lightricks | 2026-01 and 2026-03 | 19B and 22B (checkpoint file names) | Synchronized video and audio | LTX-2 Community License (2026-01-05) | Entities with annual revenue of at least 10 million US dollars must obtain a paid commercial licence; derivatives stay under the same agreement [18][19][20] |
| HunyuanVideo 1.5 | [Tencent](https://aiwiki.ai/wiki/tencent) | 2025-11-20 | 8.3B | 480p and 720p, upscaling to 1080p | Tencent Hunyuan Community License | As HunyuanVideo [12][13] |
| SkyReels V2 | Skywork AI | 2025-04-21 | 1.3B, 14B | 540P and 720P; diffusion-forcing variants for long video | Skywork Community License | Commercial use allowed under the licence terms; no military or unlawful use [51][52][53] |
| LTX-Video 0.9.6 to 0.9.8 | Lightricks | From 2025-04-15 | 2B, 13B | 30 fps at 1216x704 | LTXV Open Weights License (2025-04-15) | Entities with annual revenue of at least 10 million US dollars are directed to a paid commercial licence [14][15] |
| [HunyuanVideo](https://aiwiki.ai/wiki/hunyuan_video) | Tencent | 2024-12-03 | Over 13B | 720x1280, 129 frames | Tencent Hunyuan Community License | Does not apply in the EU, UK or South Korea; licensees above 100 million monthly active users must request a licence; outputs may not be used to improve other AI models [10][11] |
| CogVideoX-5B, 5B-I2V, 1.5-5B | Zhipu AI (zai-org) | 2024-08-27; 2024-09-19; 2024-11-08 | 5B | 720x480, 6 s, 8 fps; 1.5: 1360x768, 5 or 10 s, 16 fps | CogVideoX License | Commercial use requires registering for a basic commercial licence, valid up to 1 million visits per month [23][24][25][26] |
| [Stable Video Diffusion](https://aiwiki.ai/wiki/stable_video_diffusion) (SVD, SVD-XT) | [Stability AI](https://aiwiki.ai/wiki/stability_ai) | 2023-11-21 | Not given here | Image to video; 14 or 25 frames at 576x1024 | Stability AI Community License (last updated 2024-07-05) | Free below 1 million US dollars in annual revenue, with registration for commercial use; above that an enterprise licence is required [47][48][50] |

Notes on this group:

- MiniMax H3 derivatives: Pika H3 Flex and Hyperflow H3 are adapter weights for H3 and are distributed under the same agreement. Pika's notice file calls its adapters "Model Derivatives of MiniMax H3", and Video Rebirth's card says the weights are not licensed in the EU, UK, South Korea or US without separate authorization from MiniMax and that Hyperflow is not an official MiniMax product.[45][46]
- LTX-2.5: the GitHub repository says the code is under Apache 2.0 and that "the LTX models are not part of that".[22] Its instructions tell users who receive a 401 or 403 error to accept the model terms on Hugging Face.[22]
- SkyReels V2: the licence file in the weights repository refers to the Skywork Community License, a document whose definitions name the "Skywork Model" as the Skywork-13B model.[52][53]
- CogVideoX License: free use is granted for academic research, generated content belongs to the user to the extent local law permits, and military use is prohibited.[24]

### Research-only licences

| Model | Developer | Weights released | Basis | Documented output | Licence | Restriction |
|---|---|---|---|---|---|---|
| FastFrames-L | [Apple](https://aiwiki.ai/wiki/apple) | 2026-11-16 | 14B; lists Wan2.1-T2V-14B as base model | 832x480, 81 frames at 16 fps | Apple Machine Learning Research Model License Agreement (tagged "apple-amlr") | Research purposes only; no commercial exploitation or use in a product [54][55] |
| LTX-Video 0.9 and 0.9.1 | Lightricks | Licence dated 2024-11-22 | 2B model | Not listed here | RAIL-M License | Academic or research purposes only [16] |

## Choosing by use case

The groupings below tie each use to a capability the developer documents. They are not quality judgements.

### Consumer graphics cards

Several developers state that a variant runs on a single consumer [GPU](https://aiwiki.ai/wiki/gpu). Alibaba says Wan2.1-T2V-1.3B needs 8.19 GB of VRAM, and that Wan2.2-TI2V-5B runs on a GPU with at least 24 GB.[6][9] Tencent gives 14 GB as the minimum for HunyuanVideo 1.5 with model offloading.[12] The CogVideoX cards list Diffusers inference "starting from" 4 GB for the 2B model and 5 GB for the 5B model with the memory optimizations they describe.[25] Sand.ai lists a single RTX 4090 for MAGI-1 4.5B, and Skywork gives about 14.7 GB peak VRAM for the 1.3B SkyReels V2 model at 540P.[29][51] Lower-memory builds also exist as [quantized](https://aiwiki.ai/wiki/quantization) files: Lightricks publishes fp8 and nvfp4 versions of LTX-2, and Tencent released FP8 weights for HunyuanVideo on December 18, 2024.[10][18]

### Video with generated audio

MiniMax H3 and its two adapter derivatives generate video with synchronized stereo audio.[41][45][46] LTX-2 is described by Lightricks as an audio-video model that generates synchronized video and audio in a single model, with the caveat on its card that audio without speech may be of lower quality.[18] MAGI-2 Preview produces sound alongside the video, MOVA targets lip-synced speech and sound effects, and Ovi generates video and audio together from text or text and image.[30][35][37] Wan2.2-S2V-14B works in the other direction: it takes audio as an input and generates video driven by it.[9]

### Long clips

LongCat-Video is pretrained on video continuation, which Meituan says lets it produce minutes-long videos.[33] Skywork describes the diffusion-forcing versions of SkyReels V2 as able to generate infinite-length video.[51] MAGI-1 generates video chunk by chunk, starting the next 24-frame chunk before the current one is fully denoised.[29] Among fixed-length models, Cosmos3-Super and Cosmos3-Nano document up to 400 frames (Cosmos3-Edge 50 to 150), Step-Video-T2V up to 204 frames, and MiniMax H3 up to 15 seconds, while MAGI-2 Preview supports only 10-second clips.[30][32][38][39][41][72]

### Control and editing

Wan 2.1 VACE is presented by Alibaba as an all-in-one model for video creation and editing, and Wan2.2-Animate-14B animates or replaces a character in an existing video.[8][9] Wan 2.1 also has a first-last-frame model, released on April 17, 2025.[7] MiniMax H3's reference mode accepts up to 9 images, 3 video clips and 3 audio clips as references.[41] For LTX-2.5, Lightricks publishes separate LoRA repositories for tasks such as restoration, colorization, deblurring and clean-plate generation.[69]

## Leaderboard standing

Artificial Analysis ranks video models by Elo scores computed from blind pairwise preference votes, cast by what it describes as validated private evaluators.[1] It runs separate boards for text-to-video, image-to-video and video editing, and marks entries that have open weights.[1][2][3] The positions below were read on November 19, 2026.

| Board | What it measures | Entries | Entries flagged as open weights: position (Elo) | Board leader (Elo) |
|---|---|---|---|---|
| AA-Video-T2V v2.1 [1] | Video with audio from a text prompt | 35 | MiniMax H3 (768p) 3rd (1159); Hyperflow H3 5th (1154); Pika H3 Flex 15th (1124); LTX-2.5 Fast 29th (939); LTX-2.5 Pro 30th (938); LTX-2.3 Pro 33rd (885); LTX-2.3 Fast 34th (836) | Wan 3.0 (1174) |
| AA-Video-I2V v2.1 [2] | Video with audio from a source image | 35 | MiniMax H3 (768p) 4th (1134); Hyperflow H3 7th (1119); Pika H3 Flex 10th (1114); LTX-2.5 Pro 31st (912); LTX-2.5 Fast 32nd (912); LTX-2.3 Pro 34th (837); LTX-2.3 Fast 35th (832) | Dreamina Seedance 2.5 (1161) |
| AA-Video-Editing v1.1 [3] | Editing a video from a text instruction, audio kept | 12 | Hyperflow H3 3rd (1156); MiniMax H3 (768p) 6th (1126); Pika H3 Flex 10th (1045) | Wan 3.0 (1197) |

Three cautions apply. Arena scores move as votes accumulate and as models are added, so these positions are a snapshot. Elo values from different boards, or from different leaderboard operators, are not on a common scale and cannot be compared with each other. The boards also cover only models that Artificial Analysis has chosen to test: Wan 2.1, Wan 2.2, HunyuanVideo, Mochi 1, CogVideoX and SkyReels V2 were not listed on any of the three boards on that date.[1][2][3]

MAGI-2 Preview appeared on the text-to-video board (20th, Elo 1035) and the image-to-video board (25th, Elo 965) without an open-weights flag, although its weights are on Hugging Face under an Apache 2.0 tag.[1][2][30] Artificial Analysis also labels several entries that carry no open-weights flag as based on an open-weight model: Utopai X, HeyGen Video 1, MiniMax H3 Max and P-Video-2-Pro are marked as based on MiniMax H3, and Agnes-Video-2.5 as based on LTX-2.5.[1]

For models outside the arena, the available adoption signal is the monthly download count on Hugging Face, read on November 19, 2026. The table lists the model repositories used as sources for the rows above that showed more than 75,000 downloads. It leaves out repackaged copies and other repositories of the same families, some of which showed higher counts: Wan-AI/Wan2.1-T2V-1.3B-Diffusers showed 411,732 and Wan-AI/Wan2.2-Animate-14B 236,050.[65]

| Repository | Downloads last month |
|---|---|
| MiniMaxAI/MiniMax-H3 [41] | 5,713,769 |
| Lightricks/LTX-2.5 [70] | 2,871,190 |
| nvidia/Cosmos3-Nano [72] | 2,345,534 |
| Lightricks/LTX-Video [14] | 1,228,592 |
| Lightricks/LTX-2.3 [19] | 933,032 |
| Lightricks/LTX-2 [18] | 301,390 |
| OpenMOSS-Team/MOVA-360p [36] | 214,287 |
| stabilityai/stable-video-diffusion-img2vid-xt [47] | 202,254 |
| Pika-Labs/Pika-H3-Flex [45] | 166,236 |
| Wan-AI/Wan2.1-VACE-14B [8] | 126,131 |
| Wan-AI/Wan2.1-T2V-14B [6] | 101,839 |
| tencent/HunyuanVideo [10] | 75,263 |

Because the count depends on which files a repository contains and on how runtimes fetch them, it is a rough indicator of use and not a measure of quality.[4]

## Documented hardware requirements

The figures below are quoted from the developers' own documentation, with the conditions they state. Requirements for other resolutions, clip lengths or precisions are not given here.

| Model or variant | Stated requirement | Conditions stated |
|---|---|---|
| Wan2.1-T2V-1.3B [6] | 8.19 GB VRAM; about 4 minutes on an RTX 4090 | 5-second 480P video, without quantization |
| Wan2.2-TI2V-5B [9] | At least 24 GB VRAM (RTX 4090 named); under 9 minutes on a single consumer GPU | 5-second 720P video, without specific optimization |
| Wan2.2 T2V-A14B and I2V-A14B [9] | At least 80 GB VRAM | Single-GPU command in the repository |
| HunyuanVideo [10] | 60 GB minimum (720x1280, 129 frames); 45 GB (544x960, 129 frames); 80 GB recommended | Developer's reference pipeline |
| HunyuanVideo 1.5 [12] | 14 GB minimum | Model offloading enabled |
| CogVideoX-2B and 5B [25] | From 4 GB (2B, FP16) and from 5 GB (5B, BF16) | Diffusers with the card's memory optimizations |
| CogVideoX1.5-5B [26] | From 10 GB | Diffusers, BF16 |
| Mochi 1 preview [27] | About 60 GB on a single GPU in Genmo's repository; 42 GB or 22 GB in the Diffusers examples | 22 GB figure uses the bfloat16 variant |
| MAGI-1 [29] | 24B: 8 H100 or H800 GPUs; 24B distilled with fp8 quantization: 4 H100 or H800, or 8 RTX 4090; 4.5B: 1 RTX 4090 | Recommended machines per variant |
| MAGI-2 Preview [30] | 8 NVIDIA Hopper GPUs; about 307 GB of weights | Reference inference code |
| Step-Video-T2V [32] | 77.64 GB peak; 80 GB GPUs recommended | 544x992, 204 frames |
| MOVA [35] | 48 GB VRAM with component-wise offload; 12 GB with layerwise group offload | 8-second 360p video |
| Ovi [37] | 32 GB minimum; 24 GB with fp8 weights | Stated alongside timings for a 121-frame 720x720 video with 50 denoising steps; fp8 with slight quality degradation |
| SkyReels V2 [51] | About 14.7 GB peak (1.3B); about 51.2 GB peak (14B, diffusion forcing) | 540P video |
| Cosmos3-Super [38] | About 55 seconds per video on H200 GPUs | 50 steps, recommended 8-GPU serving configuration |
| Stable Video Diffusion [47] | About 100 s (SVD) and 180 s (SVD-XT) | A100 80GB card, default settings at release |

Genmo adds that ComfyUI can run Mochi 1 in less than 20 GB of VRAM, and ComfyUI's own documentation says the Wan 2.2 5B model should fit in 8 GB with its native offloading.[27][59] The MiniMax H3 card gives no VRAM figure; its example commands use four GPUs, and Pika reports that its 480p and 768p command-line requests passed on a single H100 with 80 GB.[41][45]

## Runtimes and their licences

The model licence covers the weights. The software used to run them carries its own licence.

| Runtime | Licence (as of November 2026) | Documented support for models in this article |
|---|---|---|
| [ComfyUI](https://aiwiki.ai/wiki/comfyui) [56] | GPL-3.0 | Wan 2.1 from February 27, 2025; Wan 2.2 from release day; official tutorials for Wan 2.2 and MiniMax H3 [7][9][59][60] |
| [Diffusers](https://aiwiki.ai/wiki/diffusers) [57] | Apache 2.0 | Pipeline pages for Wan, HunyuanVideo, HunyuanVideo 1.5, LTX-Video, LTX-2, CogVideoX, Mochi, SkyReels V2, Cosmos 3 and MiniMax H3 [58] |
| DiffSynth-Studio [64] | Apache 2.0 | Listed by Alibaba as providing video-to-video, FP8 quantization and LoRA training for Wan 2.1 [7] |

On GitHub, ComfyUI had 147,518 stars and Diffusers 34,941 as of November 2026.[56][57]

The sources on commercial use of MiniMax H3 do not say the same thing. ComfyUI's tutorial says: "Commercial use of locally generated outputs requires a MiniMax commercial license, available through Comfy, the only official reseller."[60] Comfy's licence page for H3 says the H3 Community License "generally permits commercial use", provided the licensee's commercial products and services generate no more than 20 million US dollars in yearly revenue and the other terms are met, and that a commercial licence must be acquired to use H3 in the United States, the European Union, the United Kingdom or South Korea.[75] The licence text itself grants a royalty-free licence "solely within the Applicable Territory", which excludes those four territories, and requires separate prior written authorization from MiniMax when a licensee's commercial products and services generate more than 20 million US dollars in yearly revenue.[42] The licence text does not mention Comfy or a reseller.[42]

## Limits and risks

### Restrictions on use and outputs

Most of the licences either disclaim any claim to generated outputs or say that outputs belong to the user: this is true of the Tencent, MiniMax, Stability AI, Skywork, LTXV and CogVideoX terms and of Alibaba's statement for Wan.[6][11][15][24][42][48][53] Several nonetheless restrict what may be done with outputs. The Tencent and MiniMax agreements forbid using the models or their outputs to improve other AI models, and both forbid using the works or their outputs outside the territory the licence covers.[11][42] The Stability AI licence, which says the user owns outputs, also bars using outputs to create or improve a foundational generative AI model.[48] The LTX-2 agreement applies its use restrictions to outputs as well as to the model, and treats models trained on outputs as derivatives that must stay under the same agreement.[20]

MiniMax's licence Q&A says the territorial limit reflects AI regulation in the excluded regions and means "not yet", not "not ever".[43]

### Safety filters, watermarking and provenance

Open weights can be run without a hosted service's filters, so the safeguards that ship with them vary:

- Stable Video Diffusion's released inference code enables image-level [watermarking](https://aiwiki.ai/wiki/watermarking) by default through the imWatermark library.[47]
- Lightricks publishes an LTX-2 Marking Pack for opt-in invisible marking of video frames and audio. Its card says the marks are not cryptographic signatures and that the absence of a mark does not prove a clip was not generated.[61] The LTX-2.x licence forbids removing, disabling or circumventing safety measures, metadata, watermarking and content provenance features.[21]
- The MiniMax H3 licence requires anyone offering a generation service built on H3 to maintain safeguards and not to disable them, and encourages licensees to add an AI-generation identifier to output files.[42] The model card's "Safety Guardrails" passage, which describes automated moderation of submitted text, images and videos, appears in its description of H3-Context-IR, a hosted component that the card says is not included in the weights release.[41]
- Tencent's acceptable use policy for HunyuanVideo prohibits overriding or circumventing its safety guardrails.[11]
- The Cosmos 3 card's request examples enable a guardrails option and refer to a separate `cosmos-guardrail` package.[38]

### Licence changes and withheld components

Licences in this field have changed between versions of the same family, in both directions.

| Family | Change | Date |
|---|---|---|
| LTX-Video and LTX-2 | RAIL-M, research only (0.9, 0.9.1) [16]; Open RAIL-M, which states that it allows commercial use (0.9.5) [17]; LTXV Open Weights License (0.9.6 onward) [15]; LTX-2 Community License [20]; LTX-2.x Community License [21] | 2024-11-22; 2025-03-05; 2025-04-15; 2026-01-05; 2026-08-11 (amended 2026-10-29) |
| CogVideoX-2B | Licence changed to Apache 2.0 [23] | Announced 2024-08-27 |
| Stable Video Diffusion | Launched as a research preview that Stability AI said was "not intended for real-world or commercial applications at this stage" [49][68]; weights now carry the Stability AI Community License [48] | 2023-11-21; licence last updated 2024-07-05 |

Some releases are partial. MiniMax states that its H3-Regenerate-2K module, which produces 2K output, is not yet released, so local runs produce 768p, and that H3-Context-IR, the system that prepares inputs for the generator, is not included in the release and is offered through an API.[41] Pika offers 1080p, 2K and 4K output only through its hosted service.[45] Later members of a family may also not be released as weights: on November 19, 2026 the newest repositories in the Wan-AI organization on Hugging Face were Wan 2.2 and Wan-Dancer models, while the Wan 3.0 and Wan 2.7 entries on Artificial Analysis carried no open-weights flag, and the same was true of SkyReels V4.[1][3][65]

### Copyright litigation

On September 16, 2025, Disney, Universal and Warner Bros. Discovery filed a copyright lawsuit against MiniMax in a district court in California over [Hailuo](https://aiwiki.ai/wiki/hailuo) AI, its image and video generation service, alleging that it generated copyrighted characters such as Darth Vader, the Minions and Wonder Woman, according to Reuters.[62] On May 22, 2026 the U.S. District Court for the Central District of California denied the defendants' motion to dismiss, finding personal jurisdiction and holding that the plaintiffs had plausibly alleged infringement both in training and in generated outputs, according to a summary published by the law firm Loeb & Loeb.[63] The summary names two defendants, SXJT of Shanghai and Nanonoble of Singapore, as acting under the name MiniMax, and the H3 licence defines MiniMax as Nanonoble Pte. Ltd.[42][63] MiniMax's licence Q&A for H3 says the company "is also involved in ongoing copyright-related legal proceedings specifically concerning generative video AI".[43] The complaint concerns the hosted Hailuo service. Neither source reports a final judgment.

## How open-weight models compare with closed models

On the Artificial Analysis text-to-video board read on November 19, 2026, the two entries above MiniMax H3 (768p) were Alibaba's Wan 3.0 (Elo 1174) and Utopai X (Elo 1168), neither of which is flagged as open weights.[1] The intervals shown for Wan 3.0 (1161 to 1187) and MiniMax H3 (1146 to 1172) overlap.[1] On the same board [Dreamina Seedance 2.5](https://aiwiki.ai/wiki/seedance_2_5) stood at 1156 and [Veo 3.1](https://aiwiki.ai/wiki/veo_3_1) at 963, while the LTX-2.5 and LTX-2.3 entries stood between 836 and 939.[1] On the editing board, the first open-weight entry was Hyperflow H3 at 1156, against 1197 for Wan 3.0.[3]

Apart from MAGI-2 Preview, none of the models in the permissive tier of this article was listed on these boards on that date.[1][2][3]

## See also

- [List of open-weight models](https://aiwiki.ai/wiki/list_of_open_weight_models)
- [Open-weight license comparison](https://aiwiki.ai/wiki/open_weight_license_comparison)
- [Best AI video generators](https://aiwiki.ai/wiki/best_ai_video_generators)
- [Best open source LLMs](https://aiwiki.ai/wiki/best_open_source_llms)
- [Best local LLMs](https://aiwiki.ai/wiki/best_local_llms)
- [Local AI](https://aiwiki.ai/wiki/local_ai)
- [Open source license](https://aiwiki.ai/wiki/open_source_license)

## References

1. Artificial Analysis. "AA-Video-T2V v2.1 Leaderboard - Top AI Video Models". Artificial Analysis, accessed 2026-11-19. https://artificialanalysis.ai/video/leaderboard/text-to-video
2. Artificial Analysis. "AA-Video-I2V v2.1 Leaderboard - Top AI Video Models". Artificial Analysis, accessed 2026-11-19. https://artificialanalysis.ai/video/leaderboard/image-to-video
3. Artificial Analysis. "AA-Video-Editing v1.1 Leaderboard - Top AI Video Models". Artificial Analysis, accessed 2026-11-19. https://artificialanalysis.ai/video/leaderboard/video-editing
4. Hugging Face. "Models Download Stats". Hugging Face Hub documentation, accessed 2026-11-19. https://huggingface.co/docs/hub/models-download-stats
5. Hugging Face. "Gated models". Hugging Face Hub documentation, accessed 2026-11-19. https://huggingface.co/docs/hub/models-gated
6. Wan-AI. "Wan2.1-T2V-14B" (model card). Hugging Face, accessed 2026-11-19. https://huggingface.co/Wan-AI/Wan2.1-T2V-14B
7. Wan-Video. "Wan2.1" (repository README). GitHub, accessed 2026-11-19. https://github.com/Wan-Video/Wan2.1
8. Wan-AI. "Wan2.1-VACE-14B" (model card). Hugging Face, accessed 2026-11-19. https://huggingface.co/Wan-AI/Wan2.1-VACE-14B
9. Wan-Video. "Wan2.2" (repository README). GitHub, accessed 2026-11-19. https://github.com/Wan-Video/Wan2.2
10. Tencent. "HunyuanVideo" (model card). Hugging Face, accessed 2026-11-19. https://huggingface.co/tencent/HunyuanVideo
11. Tencent. "Tencent Hunyuan Community License Agreement" (HunyuanVideo, release date December 3, 2024). Hugging Face. https://huggingface.co/tencent/HunyuanVideo/blob/main/LICENSE
12. Tencent. "HunyuanVideo-1.5" (model card). Hugging Face, accessed 2026-11-19. https://huggingface.co/tencent/HunyuanVideo-1.5
13. Tencent. "Tencent Hunyuan Community License Agreement" (HunyuanVideo 1.5, release date November 21, 2025). Hugging Face. https://huggingface.co/tencent/HunyuanVideo-1.5/blob/main/LICENSE
14. Lightricks. "LTX-Video Model Card". Hugging Face, accessed 2026-11-19. https://huggingface.co/Lightricks/LTX-Video
15. Lightricks. "LTXV Open Weights License 0.X". Hugging Face, license date April 15, 2025. https://huggingface.co/Lightricks/LTX-Video/blob/main/LTX-Video-Open-Weights-License-0.X.txt
16. Lightricks. "LTX Video 0.9 RAIL-M License". Hugging Face, dated November 22, 2024. https://huggingface.co/Lightricks/LTX-Video/blob/main/ltx-video-2b-v0.9.license.txt
17. Lightricks. "LTX Video 2B 0.9.5 Open RAIL-M License". Hugging Face, dated March 5, 2025. https://huggingface.co/Lightricks/LTX-Video/blob/main/ltx-video-2b-v0.9.5.license.txt
18. Lightricks. "LTX-2" (model card). Hugging Face, accessed 2026-11-19. https://huggingface.co/Lightricks/LTX-2
19. Lightricks. "LTX-2.3" (model card). Hugging Face, accessed 2026-11-19. https://huggingface.co/Lightricks/LTX-2.3
20. Lightricks. "LTX-2 Community License Agreement". GitHub, license date January 5, 2026. https://github.com/Lightricks/LTX-2/blob/main/LICENSE-2
21. Lightricks. "LTX-2.x Community License Agreement". GitHub, license date August 11, 2026, amended October 29, 2026. https://github.com/Lightricks/LTX-2/blob/main/LICENSE-2_x
22. Lightricks. "LTX-2" (repository README and changelog). GitHub, accessed 2026-11-19. https://github.com/Lightricks/LTX-2
23. zai-org. "CogVideo" (repository README). GitHub, accessed 2026-11-19. https://github.com/zai-org/CogVideo
24. zai-org. "The CogVideoX License". Hugging Face, accessed 2026-11-19. https://huggingface.co/zai-org/CogVideoX-5b/blob/main/LICENSE
25. zai-org. "CogVideoX-5b" (model card). Hugging Face, accessed 2026-11-19. https://huggingface.co/zai-org/CogVideoX-5b
26. zai-org. "CogVideoX1.5-5B" (model card). Hugging Face, accessed 2026-11-19. https://huggingface.co/zai-org/CogVideoX1.5-5B
27. Genmo. "mochi-1-preview" (model card). Hugging Face, accessed 2026-11-19. https://huggingface.co/genmo/mochi-1-preview
28. hpcaitech. "Open-Sora" (repository README). GitHub, accessed 2026-11-19. https://github.com/hpcaitech/Open-Sora
29. Sand AI. "MAGI-1" (model card). Hugging Face, accessed 2026-11-19. https://huggingface.co/sand-ai/MAGI-1
30. Sand AI. "MAGI-2-preview" (model card). Hugging Face, accessed 2026-11-19. https://huggingface.co/sand-ai/MAGI-2-preview
31. Sand.ai. "MAGI-2 Preview: Scaling Video Generation Models Efficiently". Sand.ai, August 5, 2026. https://sand.ai/blog/magi-2-preview
32. StepFun. "stepvideo-t2v" (model card). Hugging Face, accessed 2026-11-19. https://huggingface.co/stepfun-ai/stepvideo-t2v
33. meituan-longcat. "LongCat-Video" (model card). Hugging Face, accessed 2026-11-19. https://huggingface.co/meituan-longcat/LongCat-Video
34. meituan-longcat. "LongCat-Video" (repository README). GitHub, accessed 2026-11-19. https://github.com/meituan-longcat/LongCat-Video
35. OpenMOSS. "MOVA" (repository README). GitHub, accessed 2026-11-19. https://github.com/OpenMOSS/MOVA
36. OpenMOSS-Team. "MOVA-360p" (model card). Hugging Face, accessed 2026-11-19. https://huggingface.co/OpenMOSS-Team/MOVA-360p
37. Low, Chetwin; Wang, Weimin; Katyal, Calder. "Ovi: Twin Backbone Cross-Modal Fusion for Audio-Video Generation" (repository README). GitHub, accessed 2026-11-19. https://github.com/character-ai/Ovi
38. NVIDIA. "Cosmos 3: Omnimodal World Models for Physical AI" (Cosmos3-Super model card). Hugging Face, accessed 2026-11-19. https://huggingface.co/nvidia/Cosmos3-Super
39. NVIDIA. "Cosmos3-Edge" (model card). Hugging Face, accessed 2026-11-19. https://huggingface.co/nvidia/Cosmos3-Edge
40. OpenMDW. "OpenMDW-1.1". openmdw.ai, accessed 2026-11-19. https://openmdw.ai/license/1-1/
41. MiniMax. "MiniMax-H3" (model card). Hugging Face, accessed 2026-11-19. https://huggingface.co/MiniMaxAI/MiniMax-H3
42. MiniMax. "MiniMax H3 Community License Agreement". Hugging Face, license date August 2, 2026. https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE
43. MiniMax. "Q&A About License". Hugging Face, accessed 2026-11-19. https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/docs/QA-about-License.md
44. MiniMax. "Open General Intelligence: MiniMax H3 Is Now Open Source". MiniMax News, August 3, 2026. https://www.minimax.io/news/minimax-h3-open-source
45. Pika Labs. "Pika-H3-Flex" (model card and NOTICE file). Hugging Face, accessed 2026-11-19. https://huggingface.co/Pika-Labs/Pika-H3-Flex
46. Video Rebirth. "hyperflow" (model card). Hugging Face, accessed 2026-11-19. https://huggingface.co/VideoRebirthAI/hyperflow
47. Stability AI. "stable-video-diffusion-img2vid-xt" (model card). Hugging Face, accessed 2026-11-19. https://huggingface.co/stabilityai/stable-video-diffusion-img2vid-xt
48. Stability AI. "Stability AI Community License Agreement". Hugging Face, last updated July 5, 2024. https://huggingface.co/stabilityai/stable-video-diffusion-img2vid-xt/blob/main/LICENSE.md
49. Stability AI. "Introducing Stable Video Diffusion". Stability AI, November 21, 2023. https://stability.ai/news-updates/stable-video-diffusion-open-ai-video-model
50. Stability AI. "License Update". Stability AI, July 5, 2024. https://stability.ai/news-updates/license-update
51. SkyworkAI. "SkyReels V2: Infinite-Length Film Generative Model" (repository README). GitHub, accessed 2026-11-19. https://github.com/SkyworkAI/SkyReels-V2
52. Skywork. "SkyReels-V2-DF-14B-540P" (LICENSE file). Hugging Face, accessed 2026-11-19. https://huggingface.co/Skywork/SkyReels-V2-DF-14B-540P/blob/main/LICENSE
53. Skywork. "Skywork Community License". GitHub, accessed 2026-11-19. https://github.com/SkyworkAI/Skywork/blob/main/Skywork%20Community%20License.pdf
54. Apple. "FastFrames-L" (model card). Hugging Face, accessed 2026-11-19. https://huggingface.co/apple/FastFrames-L
55. Apple. "FastFrames-L" (LICENSE file). Hugging Face, accessed 2026-11-19. https://huggingface.co/apple/FastFrames-L/blob/main/LICENSE
56. Comfy-Org. "ComfyUI" (repository and LICENSE). GitHub, as of November 2026. https://github.com/Comfy-Org/ComfyUI
57. Hugging Face. "diffusers" (repository and LICENSE). GitHub, as of November 2026. https://github.com/huggingface/diffusers
58. Hugging Face. "Pipelines". Diffusers documentation, accessed 2026-11-19. https://huggingface.co/docs/diffusers/main/en/api/pipelines/overview
59. ComfyUI. "Wan2.2 Video Generation ComfyUI Official Native Workflow Example". ComfyUI documentation, accessed 2026-11-19. https://docs.comfy.org/tutorials/video/wan/wan2_2
60. ComfyUI. "ComfyUI MiniMax H3 Video Generation Guide". ComfyUI documentation, accessed 2026-11-19. https://docs.comfy.org/tutorials/video/minimax/minimax-h3
61. Lightricks. "LTX-2 Marking Pack (VideoSeal / AudioSeal)" (model card). Hugging Face, accessed 2026-11-19. https://huggingface.co/Lightricks/LTX-2-Marking
62. Reuters. "Disney, Universal, Warner Bros Discovery sue China's MiniMax for copyright infringement". The Standard, September 16, 2025. https://www.thestandard.com.hk/innovation/article/311715/Disney-Universal-Warner-Bros-Discovery-sue-Chinas-MiniMax-for-copyright-infringement
63. Dickstein, Tal; Sievers, Jacob M. "Disney Enterprises, Inc. v. Minimax". Loeb & Loeb LLP, June 2026. https://www.loeb.com/en/insights/publications/2026/06/disney-enterprises-inc-v-minimax
64. ModelScope. "DiffSynth-Studio" (repository and LICENSE). GitHub, as of November 2026. https://github.com/modelscope/DiffSynth-Studio
65. Wan-AI. Organization page and model list. Hugging Face, accessed 2026-11-19. https://huggingface.co/Wan-AI
66. Stability AI. "stable-audio-open-1.0" (repository metadata and public page; licence text gated). Hugging Face, accessed 2026-11-19. https://huggingface.co/stabilityai/stable-audio-open-1.0
67. Comfy-Org. "Wan-2.2" (repository). Hugging Face, accessed 2026-11-19. https://huggingface.co/Comfy-Org/Wan-2.2
68. Stability AI. "generative-models" (repository README). GitHub, accessed 2026-11-19. https://github.com/Stability-AI/generative-models
69. Lightricks. Organization page and model list. Hugging Face, accessed 2026-11-19. https://huggingface.co/Lightricks
70. Lightricks. "LTX-2.5" (repository metadata; model card gated). Hugging Face, accessed 2026-11-19. https://huggingface.co/Lightricks/LTX-2.5
71. Wan-AI. "Wan2.2-TI2V-5B" (model card). Hugging Face, accessed 2026-11-19. https://huggingface.co/Wan-AI/Wan2.2-TI2V-5B
72. NVIDIA. "Cosmos3-Nano" (model card). Hugging Face, accessed 2026-11-19. https://huggingface.co/nvidia/Cosmos3-Nano
73. chetwinlow1. "Ovi" (weights repository). Hugging Face, accessed 2026-11-19. https://huggingface.co/chetwinlow1/Ovi
74. hpcai-tech. "Open-Sora-v2" (weights repository). Hugging Face, accessed 2026-11-19. https://huggingface.co/hpcai-tech/Open-Sora-v2
75. Comfy. "MiniMax Commercial License on Comfy". comfy.org, updated August 2026, accessed 2026-11-19. https://comfy.org/minimax/license
76. Wan-AI. "Wan-Dancer-14B" (model card). Hugging Face, accessed 2026-11-19. https://huggingface.co/Wan-AI/Wan-Dancer-14B
77. Wan-AI. "Wan2.2-Animate-2-14B" (model card). Hugging Face, accessed 2026-11-19. https://huggingface.co/Wan-AI/Wan2.2-Animate-2-14B

