Are the MiniMax H3 weights actually downloadable?

Yes. MiniMax H3 open weights landed on Hugging Face at MiniMaxAI/MiniMax-H3 on Aug 2, 2026 under the MiniMax H3 Community License. The full model is 33B parameters, and roughly 13B of AdaLN branches can be skipped at inference. ComfyUI shipped day-0 support and quantized checkpoints for local use.

What can MiniMax H3 do besides text-to-video?

MiniMax H3 unifies text-to-image, text-to-video, text-to-audio, video editing, and motion transfer in one model — no task-specific model swap. A single request can mix up to nine reference images, three video clips, and three audio tracks, letting users transfer motion from a reference clip or restyle an existing video from a prompt.

How does MiniMax H3 compare on price?

MiniMax says H3's per-second 2K price is under a third of what mainstream rivals charge, and its 768p tier is under half the cost of competing 720p offerings. That aligns with the company's earlier LLM strategy of undercutting Chinese and Western frontier peers on token cost.

Can I run MiniMax H3 on a consumer GPU?

With ComfyUI's day-0 support, MiniMax H3 can run on a single RTX 3060. ComfyUI pruned roughly 40% of the modulation weights and quantized to fp8/int8, dropping the memory footprint from 123.6 GB to 42.5 GB and enabling local inference on consumer hardware, though generation is slow at that end of the range.

How is MiniMax H3 doing this at 2K in one model?

MiniMax H3 uses a new H3-VAE tokenizer that MiniMax says gives a 4× gain in effective sequence length, and compresses long contextual prompts from roughly 100K tokens down to about 4K. That lets one model handle 15-second 2K generations with stereo audio without splitting into separate stages.

MiniMax · 2026-07-31 · major

MiniMax H3 — open-weights video model does 2K, 15s, and native stereo sound

MiniMax H3 generates 15-second 2K clips with synchronized stereo audio in one model. Open weights are now live on Hugging Face under the MiniMax Community License, and ComfyUI shipped day-0 support for local inference.

A single open-weights video model does 2K clips, stereo audio, editing, and motion transfer up to 15 seconds long.

Quick facts

Maker	MiniMax (Shanghai)
Model	H3 (Hailuo 3.0), full-modal video
Parameters	33B (about 13B AdaLN branches skippable for inference)
Output	Up to 15s clips, native 2K (2560×1440), 24 FPS with stereo audio
License	MiniMax H3 Community License
Availability	MiniMax API, hosted providers, and open weights on Hugging Face (Aug 2)
Price vs peers	Under 1/3 of mainstream 2K video pricing

What is it?

MiniMax H3 (Hailuo 3.0) is a full-modal video generation model from the Shanghai lab MiniMax. It generates up to 15-second video at native 2K resolution with synchronized stereo audio, and unifies text-to-image, text-to-video, text-to-audio, video editing, and motion transfer inside one 33B-parameter model.

How does it work?

The new H3-VAE tokenizer gives MiniMax roughly a 4× gain in effective sequence length, and long contextual prompts compress from ~100K tokens down to ~4K. A single H3 request accepts up to nine reference images, three video clips, and three audio tracks as context, so users can restyle a clip, transfer motion from a reference, or generate voice and sound effects in one call.

Why does it matter?

Open, weight-available video generation at 2K with native audio has not existed before at this quality level — most rivals either cap out at 720p, drop audio, or lock the weights behind a proprietary API. With weights now live on Hugging Face under the MiniMax Community License and ComfyUI's day-0 pruned build fitting on an RTX 3060, MiniMax puts real pressure on Runway, Kling, and Sora clones.

Who is it for?

creators, ad agencies, and video-app builders wanting open 2K video with audio

Frequently asked questions

Are the MiniMax H3 weights actually downloadable?: Yes. MiniMax H3 open weights landed on Hugging Face at MiniMaxAI/MiniMax-H3 on Aug 2, 2026 under the MiniMax H3 Community License. The full model is 33B parameters, and roughly 13B of AdaLN branches can be skipped at inference. ComfyUI shipped day-0 support and quantized checkpoints for local use.
What can MiniMax H3 do besides text-to-video?: MiniMax H3 unifies text-to-image, text-to-video, text-to-audio, video editing, and motion transfer in one model — no task-specific model swap. A single request can mix up to nine reference images, three video clips, and three audio tracks, letting users transfer motion from a reference clip or restyle an existing video from a prompt.
How does MiniMax H3 compare on price?: MiniMax says H3's per-second 2K price is under a third of what mainstream rivals charge, and its 768p tier is under half the cost of competing 720p offerings. That aligns with the company's earlier LLM strategy of undercutting Chinese and Western frontier peers on token cost.
Can I run MiniMax H3 on a consumer GPU?: With ComfyUI's day-0 support, MiniMax H3 can run on a single RTX 3060. ComfyUI pruned roughly 40% of the modulation weights and quantized to fp8/int8, dropping the memory footprint from 123.6 GB to 42.5 GB and enabling local inference on consumer hardware, though generation is slow at that end of the range.
How is MiniMax H3 doing this at 2K in one model?: MiniMax H3 uses a new H3-VAE tokenizer that MiniMax says gives a 4× gain in effective sequence length, and compresses long contextual prompts from roughly 100K tokens down to about 4K. That lets one model handle 15-second 2K generations with stereo audio without splitting into separate stages.

Try it

huggingface.co/MiniMaxAI/MiniMax-H3 (open weights) or hailuoai.video (hosted)