HomeAIAI ModelsHeyGen's Video Model, Built on MiniMax H3, Now Makes 30-Second Clips

HeyGen’s Video Model, Built on MiniMax H3, Now Makes 30-Second Clips

HeyGen’s first general-purpose video model can now make text-to-video and image-to-video clips of up to 30 seconds, each with its own soundtrack, the company said on 23 October. HeyGen calls it the number-one text-to-video model. The public Artificial Analysis leaderboard placed it fourth when TSN checked on 28 October.

The model, HeyGen Video 1, is not built from scratch. HeyGen’s own developer documentation describes it as “built on MiniMax H3 and post-trained by HeyGen”, and the company said the same when it first released the model on 30 September. It reached HeyGen’s own app on 21 October, when the company said it was “Live on HeyGen + API” and could produce “a 2K shot with sound in under a minute”. The 30-second limit came two days later, announced by HeyGen and its developer account.

The model writes picture and sound from one prompt

HeyGen, based in Los Angeles, made its name with AI avatars that read a script on camera. Video 1 is different: it invents the scene. According to the documentation, “room tone, effects and speech are generated with the image from the same prompt”, and the model follows a brief literally: “If you did not ask for dialogue, you do not get dialogue.”

There are three modes. Text to video works from a prompt alone, image to video uses a supplied picture as the first frame, and reference to video builds a new scene that keeps products, people or places from the user’s images, video or audio. The longer 30-second clips apply to the first two modes “with compatible settings”; reference-to-video clips still run from 5 to 15 seconds.

Output is MP4 at 24 frames a second with stereo audio, at 480p, 768p, 1080p or 2K, in six aspect ratios from 21:9 to 9:16. Prompts can run to 32,000 characters, and a seed setting lets users regenerate the same clip while changing one detail at a time.

HeyGen is also clear about the limits. The documentation says the model is “at its best on short, contained shots: one subject, one place, one action, a locked or barely moving camera”, and “least reliable on long on-screen text, soft organic motion such as petals, paper and hair, and close hand work like assembling or operating equipment”.

The launch discount ends in November

HeyGen bills per second of output. Its help centre article, last updated on 26 October, lists a 50% launch discount through the end of October: $0.01 a second for text or image to video at 480p and $0.018 at 768p. From November the standard rates are $0.02 and $0.03. Reference to video goes to $0.04 and $0.06, and bills the seconds of any input video as well as the output. Clips longer than 15 seconds carry a higher promotional rate, from $0.045 to $0.105 a second. The model is available through HeyGen’s API and its Creative Studio, and through OpenRouter, which HeyGen calls its “primary launch partner”, Runware and ComfyUI.

A leaderboard sits underneath HeyGen’s number-one claim

HeyGen’s ranking claim rests on its own test. The help article says Video 1 “ranked #1 in quality” against Seedance 2.5, Seedance 2.0, H3 Max and Kling 3.0, based on 4,800 blind votes on the Artificial Analysis Arena query set, and the company repeated on X that it ranked first in a blind test on Artificial Analysis. The public leaderboard, built from 5,714 samples for this model, tells a narrower story. On 28 October HeyGen Video 1 had an Elo score of 1157, behind Wan 3.0 at 1174, Utopai X at 1168 and MiniMax’s own H3 at 768p, the base it was built on, at 1158. Those gaps are within the leaderboard’s stated margin, which puts HeyGen’s possible rank anywhere from second to seventh. “Among the leaders” is fair. “Number one” depends on which models HeyGen chose to compare.

HeyGen is selling price and packaging more than a new model

The help article describes Video 1 as “built in-house”, which sits awkwardly with HeyGen’s own documentation and launch post naming MiniMax H3 as the base. The better reading is that HeyGen’s contribution is post-training, a large prompt guide and a pricing model that undercuts rivals for now. That is a real business, and the sound-with-picture approach is the same bet behind start-ups such as Mirelo. But it means Video 1’s quality ceiling is tied to MiniMax’s, as research such as FastH3-Omni is already speeding up the same base. The test comes in November, when the launch discount ends and Video 1 competes on quality alone with models such as fal’s FLUX 3 Max, which sits two points behind it on the same table.

Sources

Related stories

Follow @tsnmediahub on X for breaking AI news as it happens.

Follow on X

Share this story

Latest stories

More in this category

Latest stories

Explore the TSN Atlas: Every topic we cover, mapped. Click a bubble to dive in.