Video model capability matrix

How to read it

• Each row is one model, and links to that model's page. Where a vendor documents several versions on one page, the row records only what the page names for that version.
• A no means a page listing what the model accepts or returns did not include it. A not established means we found no statement in either direction. The two are different and the table keeps them apart.
• Durations are the vendor's figure for a single generation. Most of these models can be extended past it, and an extended clip is several generations joined, not one long one.
• The table measures what a model exposes, not how good its output looks. Use it to rule models out on a hard requirement, then judge the picture.

What the three values mean

Not established is not a no. It records that we could not find a vendor statement either way, and never that the model lacks the feature.

What falls out of it

Sound is standard at the top of the hosted market and rare in the open

20 of the 68 video models return audio with the picture, and almost all of them are hosted: Veo 3.1 and its Fast and Lite tiers, Veo 3, Sora 2 and Sora 2 Pro, Kling 2.6 and the three Kling 3.0 tiers, the three Vidu Q3 tiers, PixVerse V6 and C1, Seedance 2.0, Higgsfield Cinematic Studio and Google Flow. Of the 22 models with published weights, only LTX-2.5 and NVIDIA Cosmos 3 do it, and on Cosmos only the Super and Nano variants list sound as an output. If a job needs dialogue or effects from the model itself, that is the shortlist.

The negative prompt has moved to the models you run yourself

20 of the 22 open weight models expose a negative prompt. Among hosted models only 11 do, and 4 of those are Pika. 28 hosted models publish a request body with no such field, including every current Runway model, Sora 2, Luma, Vidu, MiniMax, PixVerse, Adobe Firefly, Amazon Nova Reel and both Seedance generations. Kling still accepts one, but only on its older endpoints: the current ones tell you to put positive and negative description in the same prompt text. Where there is no field, exclusions have to sit in the prompt, competing for length and read as description of the thing you are trying to avoid.

Audio is often a switch, and sometimes it is priced

Veo 3.1 and Sora 2 generate audio unconditionally: Google lists it as always on. Kling 2.6 and 3.0 take a setting with the values native and off, Kling 3.0 Omni adds original for keeping a reference clip's sound, Vidu, Higgsfield and NVIDIA Cosmos 3 have a switch that is off by default, and PixVerse charges less for a silent generation than for one with audio. Native audio being available is not the same as it being on.

Tiers inside one family do not match each other

Kling 3.0 Turbo accepts a prompt and a first frame and nothing else, while Kling 3.0 and 3.0 Omni take a last frame as well. Kling O1 has an audio setting with no value for generating any. Vidu Q3 has an audio switch and Vidu Q2 does not. MiniMax Hailuo offers 1080p at 6 seconds but drops to 768p at 10. Reading a capability off the family name is how a shot gets promised and then cannot be delivered.

The longest single generation is 2 minutes, and it arrives in 6 second pieces

Amazon Nova Reel 1.1 generates up to two minutes, in six second increments, and anything over six seconds needs its 1.1 model ID. FramePack reaches 60 seconds on your own hardware. Runway Aleph 2.0 and Act-Two work to 30 seconds on a clip you supply. Straight text to video is much shorter: Sora 2 at 20 seconds, Vidu Q3 at 16, Kling 3.0 at 15, Veo 3.1 at 8.

What each column means

Native audio

The model returns a soundtrack with the picture, generated in the same pass, rather than leaving you to add sound afterwards.

The output modalities recorded for each model on 2026-08-14, checked against vendor documentation on 2026-08-17.

Negative prompt field

A separate field for what the model must leave out. Where there is none, exclusions have to sit inside the prompt itself, where they compete for length and are read as description.

The request parameters named in each vendor's own API reference, or the inference arguments named in the repository for models you run yourself.

Image as start frame

The generation can be anchored to a still you supply, which is how you control composition, cast and colour instead of describing them.

The input modalities and controls recorded for each model, checked against vendor documentation.

Last frame or keyframes

The generation can be given an end frame, or frames at set points, so a shot lands somewhere specific. This is what makes a cut between two shots match.

The request parameters named in each vendor's own API reference, or the controls recorded for each model.

Longest single generation

Seconds in one call, before extension. Longer clips are usually assembled from several calls, and the joins are visible.

The figure the vendor published on the date the model record was checked.

Open weights

The weights are published, so the model can run on your own hardware. It does not mean you may use the output commercially. Read the licence.

The model record.

API

The vendor offers a first-party programmatic endpoint. A model marked no may still be reachable through a reseller.

The model record.

Method

Every row starts from the model's own record in this directory, which was read from vendor documentation on 2026-08-14. Each capability column was then checked against the vendor's own request schema, API reference or repository on 2026-08-17, and the URL is recorded per claim. A yes names a field or an argument. A no means a page that enumerates what the model accepts or returns does not include it. Anything else is not established. Reseller schemas were not used as evidence, and where two sources disagree the disagreement is recorded on the row instead of being resolved quietly.

Common questions

Which AI video models generate sound?

20 of the 68 video models here do. Veo 3.1 and its Fast and Lite tiers, Veo 3, Sora 2 and Sora 2 Pro, Kling 2.6, Kling 3.0, Kling 3.0 Turbo and Kling 3.0 Omni, Vidu Q3-pro, Q3-turbo and Q3-mix, PixVerse V6 and C1, ByteDance Seedance 2.0, Higgsfield Cinematic Studio, Google Flow, and two models with published weights, LTX-2.5 and NVIDIA Cosmos 3. Kling 2.5 Turbo and Kling O1 do not, and neither does Seedance 1.0 Pro, so a family name is not enough to go on. Meta Movie Gen is a near miss worth knowing about: sound comes from a separate model that works on a finished video, so the video model itself does not generate it.

Which video models support a negative prompt?

Among hosted models, 11 do. Pika 2.5, Pika 2.5 Keyframe, Pikadditions and Pikaswaps all take a negative_prompt. Veo 3.1, its Fast and Lite tiers and Veo 3 have a negativePrompt in Vertex AI's request, though Google's Gemini API page for the same models does not list it. Kling keeps one on its older endpoints, which still accept Kling 2.5 Turbo, 2.6 and 3.0, while its current endpoints have dropped it. The other 28 hosted models we could check publish request bodies with no such field, including all four Runway models, Sora 2, Luma, Vidu, MiniMax, PixVerse, Adobe Firefly, Amazon Nova Reel and both Seedance generations. Models you run yourself are the opposite: 20 of 22 expose one, though often as an argument in the Python pipeline rather than a flag on the command line script. The exceptions are Stable Video Diffusion, which takes no text at all, and HunyuanVideo-Avatar, where the argument is parsed and then ignored.

Which AI video model makes the longest clip?

Amazon Nova Reel 1.1, at up to two minutes. It generates in six second increments, and clips longer than six seconds require its 1.1 model ID, so two minutes is a multi-shot sequence rather than one continuous take. FramePack reaches 60 seconds and runs on your own hardware. Runway Aleph 2.0 and Act-Two go to 30 seconds but both work on footage you supply. For plain text to video the ceiling is much lower: Sora 2 at 20 seconds, the Vidu Q3 tiers at 16, Kling 3.0 at 15, Veo 3.1 at 8.

Which AI video models accept a last frame or keyframes?

33 of the 68. Luma Ray3.2 takes up to 64 guide frames, though not in combination with a start and end frame and not at its 10 second duration. Pika 2.5 Keyframe animates between 2 and 5 ordered frames. Runway Aleph 2.0 takes up to 5 timed guidance images placed at absolute timestamps in the clip you supply. Veo 3.1 and Veo 3.1 Fast take a lastFrame, and so do Kling 3.0, 3.0 Omni and O1, MiniMax H3, PixVerse V6 and C1, Vidu Q1, Q3-pro and Q3-turbo, both Seedance generations and Adobe Firefly, which positions an image on the timeline from 0 to 1. Among open weights, LTX-Video and LTX-2.5 condition images at named frame indexes, SkyReels-V2 has an end image on its diffusion forcing checkpoints, Open-Sora 2.0 can target the tail or both ends, and Wan 2.1 VACE documents first, last and first-and-last frame modes. Runway's own Gen-4 models do not: their image position value is first only.

Which AI video models can I run on my own computer?

22 of the 68 publish weights: Wan 2.1 T2V and VACE, Wan 2.2 T2V, I2V, TI2V and S2V, Wan-Animate-2, HunyuanVideo 13B, HunyuanVideo 1.5, HunyuanVideo-I2V, HunyuanVideo-Avatar, LTX-Video 13B, LTX-2.5, CogVideoX1.5-5B, Mochi 1, SkyReels-V2, Open-Sora 2.0, Pyramid Flow, Stable Video Diffusion, Allegro, FramePack and NVIDIA Cosmos 3. Published weights are not permission to use the output commercially. Read the licence on the model's page.

Which AI video models accept an image as the first frame?

55 of the 68, which makes it the one capability you can nearly assume. The exceptions are text-only models such as Mochi 1, Allegro, HunyuanVideo 13B and the Wan text to video checkpoints, and editing models that take a clip instead, such as Runway Aleph 2.0, Pikadditions and Pikaswaps. Stable Video Diffusion is the inverse case: it takes an image and nothing else, because it cannot be steered with text at all.

What does not established mean in this table?

It means we found no statement from the vendor either way, so the cell is left empty rather than filled in. A no is different: it means a page listing what the model accepts or returns did not include the feature. Neither value is a judgement about the model, and a not established cell is as likely to become a yes as a no once the vendor documents it.

Why is a capability listed as not established when other sites state it plainly?

Because a lot of what is published about these models comes from resellers rather than from the vendor, and the two disagree. Google's Gemini API page for Veo lists the request fields without a negative prompt, while Runway, which resells Veo 3.1, publishes one on its own Veo endpoints, so this table records that as not established and says why on the row. The prompt length guide on this site records a separate negative prompt budget for PixVerse, and PixVerse's own published request bodies contain no such field. Where sources conflict the row carries the conflict.