OpenModelsHQ

Video models 35

We track 35 open video models from 5 labs, 13 of them released in the 12 months to 7 Nov 2026. The newest is LTX-2.5-22b-LoRA-Cinemagraph (10 Sep 2026) and the most downloaded is LTX-2.5, with 2,778,909 downloads in 30 days. 22 fit a 16 GB GPU at 4-bit, and none has a hosted API in our data.

VBench evaluates video-generation quality across separate dimensions, including motion and subject consistency. Its authors publish evaluation code and a leaderboard. Those measurements help distinguish problems that a single overall score can hide, but they still depend on the chosen prompts and generation settings.

What the category covers

Text-to-video is the headline, where a prompt produces a clip. Image-to-video is another supported task, taking a still and animating it, and model names carry the distinction as t2v and i2v suffixes. Some releases combine both. Beyond generation there are control-oriented variants, where motion or pose guides the output, and LoRA adapters that specialise a base model to one effect.

Resolution, frame count and frame rate must stay within the model and pipeline’s supported settings. Extending a short-clip configuration can require additional memory or a different generation method.

How they are measured

VBench evaluates defined dimensions so a model can be checked for motion quality, prompt adherence and consistency separately. The project also publishes VBench++ and VBench-2.0, with different scopes. A result from one suite should not be treated as interchangeable with a score from another. The README’s May 2025 entry adds evaluation of your own videos for VBench-2.0 specifically; the original VBench supports custom videos separately, from an entry dated March 2024, for six dimensions.

Treat lab-reported video scores with the same caution as elsewhere. A number is comparable only when the dimension, the prompt suite and the sampling settings match.

What serves them

The tooling is shared with image generation. ComfyUI’s catalogue lists video generation workflows covering model families including Wan 2.1 and 2.2, LTX-Video releases, HunyuanVideo and CogVideoX, and the node graph suits video because a pipeline typically chains several stages. Diffusers covers the library route, and many releases ship in a diffusers-ready form, which is what the -diffusers suffix on a repository name indicates.

Core vLLM and llama.cpp are token generators: vLLM’s documentation says the generative models it supports share the ability to generate text, and the llama.cpp README describes the project as LLM inference in C/C++. For other outputs, vLLM’s documentation says vLLM-Omni “supports the generation of multimodal content, including images, videos, and audio”. Use an integration that explicitly supports the video architecture; support for a lab’s text model does not establish support for its video model.

The hardware reality

Video pipelines can require substantial memory because activations depend on frame count and resolution as well as model weights. A configuration that fits an image-generation workload may not fit a video workload on the same card.

Smaller parameter variants of the same family need less memory, and reduced-precision releases cut the weight footprint further. Mixture-of-experts architectures can reduce active computation, though inactive experts still contribute to the full checkpoint size.

Generation time also depends on clip length, resolution, model architecture and sampling steps. Longer clips do not automatically change the number of denoising steps, but can increase the work performed at each step. Compare a short, low-resolution run before committing to a longer render.

Licence patterns

Read the licence attached to the exact video checkpoint. The Apache 2.0 terms, covered on our Apache 2.0 page, differ from custom lab licences. A runtime’s licence does not establish the conditions for the model it loads.

Adapters and LoRAs may have their own licences, and applicable base-model terms still need to be checked. Neither file can be assumed to override the other. Our MIT and modified MIT pages explain permissive terms and additions.

This summarises licence text. It is not legal advice.

FAQ

Can I generate video on a consumer GPU?

Smaller variants of several families are released for exactly that, and reduced-precision versions help further. Expect to work at lower resolution and shorter duration than the flagship releases demonstrate.

What does VBench actually score?

VBench scores defined aspects of video quality, with later suites extending the evaluation scope. Inspect the suite version and dimension-level results rather than relying only on an aggregate score.

What is the difference between t2v and i2v?

Text-to-video generates a clip from a prompt alone. Image-to-video animates a supplied still. Some releases do both, and the suffix in the repository name is usually the quickest way to tell.

Why do some repositories end in -diffusers?

That marks a release packaged for the Hugging Face diffusers library, which expects a particular pipeline layout. The same model may also ship in a native format for other tooling.

Pick a view

All models

Newest

ModelReleasedTypeSizeMemory at 4-bit16 GB GPUDownloads, 30 daysCheapest API, blended / 1M
10 Sep 2026VideoUnknownUnknownUnknown6,116No tracked API
26 Jul 2026Video19B13.5 GBFits32,655No tracked API
23 Jul 2026VideoUnknownUnknownUnknown2,778,909No tracked API
10 Jul 2026VideoUnknownUnknownUnknown16,897No tracked API
9 Jun 2026VideoUnknownUnknownUnknown8,181No tracked API
16 Mar 2026VideoUnknownUnknownUnknown14,760No tracked API
4 Mar 2026VideoUnknownUnknownUnknown862,819No tracked API
5 Jan 2026VideoUnknownUnknownUnknown6,634No tracked API
5 Jan 2026VideoUnknownUnknownUnknown7,836No tracked API
5 Jan 2026VideoUnknownUnknownUnknown7,631No tracked API
5 Jan 2026VideoUnknownUnknownUnknown9,158No tracked API
5 Jan 2026VideoUnknownUnknownUnknown7,545No tracked API

Most downloaded

ModelReleasedTypeSizeMemory at 4-bit16 GB GPUDownloads, 30 daysCheapest API, blended / 1M
23 Jul 2026VideoUnknownUnknownUnknown2,778,909No tracked API
31 Oct 2024Video1.9B1.4 GBFits1,180,743No tracked API
4 Mar 2026VideoUnknownUnknownUnknown862,819No tracked API
1 Mar 2025Video1.4B1 GBFits384,970No tracked API
3 Jan 2026Video18.9B13.5 GBFits297,917No tracked API
25 Feb 2025Video16.4B11.7 GBFits190,651No tracked API
28 Jul 2025Video14.3B10.2 GBFits182,320No tracked API
28 Jul 2025Video14.3B10.2 GBFits181,707No tracked API
28 Jul 2025Video5B3.6 GBFits179,371No tracked API
13 May 2025Video17.3B12.4 GBFits158,325No tracked API
20 Nov 2023Video1.5B1.1 GBFits118,085No tracked API
1 Mar 2025Video16.4B11.7 GBFits116,012No tracked API

Smallest that fit 16 GB

ModelReleasedTypeSizeMemory at 4-bit16 GB GPUDownloads, 30 daysCheapest API, blended / 1M
19 Feb 2025Video1.3B0.9 GBFits30,957No tracked API
1 Mar 2025Video1.4B1 GBFits384,970No tracked API
25 Feb 2025Video1.4B1 GBFits92,058No tracked API
20 Nov 2023Video1.5B1.1 GBFits118,085No tracked API
20 Nov 2023Video1.5B1.1 GBFits43,404No tracked API
31 Oct 2024Video1.9B1.4 GBFits1,180,743No tracked API
16 Mar 2025Video1.9B1.4 GBFits11,941No tracked API
13 May 2025Video2.2B1.5 GBFits42,200No tracked API
4 Jun 2025Video2.2B1.5 GBFits22,408No tracked API
28 Jul 2025Video5B3.6 GBFits179,371No tracked API
28 Jul 2025Video14.3B10.2 GBFits181,707No tracked API
25 Feb 2025Video14.3B10.2 GBFits115,068No tracked API
Top 12 in each view. Memory is the 4-bit estimate with an 8k context; a model fits when it stays within 90% of the GPU's memory.How we calculate

Releases per month

13 in 12 months
Video model releases per month
Months run to the newest release in our data, across all model types. Small screens show the last 6 months; the data table has all 12.
Video model releases per month: data
MonthTotal
Dec 20250
Jan 20266
Feb 20260
Mar 20262
Apr 20260
May 20260
Jun 20261
Jul 20263
Aug 20260
Sep 20261
Oct 20260
Nov 20260

Cheapest hosted APIs

All API prices

We have not recorded API pricing for these video models yet.

Top labs for video

Most tracked video models first

Sources

Other model types