LTX-2.5 or MiniMax H3: two open AI video models with sound

In August two companies published the weights of a video model that also makes the soundtrack: MiniMax opened H3 on 3 August and Lightricks released LTX-2.5 on 11 August. On paper they do the same job. The model cards show how differently.

Size

H3 is the larger: a 33-billion-parameter dense transformer with Qwen3-VL-32B as its text encoder. LTX-2.5 is a 22-billion-parameter diffusion transformer with a Gemma 4 text encoder. MiniMax's own SGLang example runs H3 on four GPUs; Lightricks ships LTX-2.5 in int8 and NVFP4 quantised versions and a distilled 8-step checkpoint, which is where it gets onto smaller hardware.

What comes out

Open H3 generates at 768p, 4 to 15 seconds at 24 fps, with 32 kHz stereo audio and dialogue in eleven languages; higher-resolution variants are hosted only. LTX-2.5 defaults to 24 fps and reaches 3840×2176 through spatial upscaling. Both cut between shots inside one clip: H3 with numbered shots in the prompt, LTX-2.5 with native multi-shot generation that keeps identities consistent across wide, medium and close shots.

References

H3's open release has two checkpoints. FL2VA works from up to two images, a first and last frame. Ref2VA takes up to nine images, three videos of 2 to 15 seconds and three audio files, twelve files in all. LTX-2.5 works from text, a first image or an audio track — the three starting points Lightricks' own API lists for it.

The licences

This is where they differ most. LTX-2.5 is under the LTX-2 Community License: free, commercial use included, for organisations under $10 million in annual revenue, with a paid agreement above that. H3's licence covers the world except the European Union, the United Kingdom, South Korea and the United States, where MiniMax asks organisations to apply for a separate licence. Elsewhere, commercial products over $20 million in yearly revenue need MiniMax's written authorisation, and any commercial product using H3 has to display "MiniMax H3" in its interface.

Which to try

If you are in one of H3's excluded territories, the question is settled before the specs: LTX-2.5, or H3 through MiniMax's hosted API. Outside them, H3 has the richer reference inputs and LTX-2.5 the quantised checkpoints for smaller hardware — and both run in ComfyUI.