MiniMax H3 vs LTX-2.5: Open Video Models Compared (2026)
MiniMax H3 and LTX-2.5 are both open-weight video models with built-in audio, but they suit different jobs. H3 is the larger model (33B parameters) with more input control: first and last frames, multiple image, video and audio references, and more aspect ratios. LTX-2.5 (22B) is much faster on the same GPU, generates up to 20 seconds, or up to 4K, through its API, and supports multi-shot clips. The biggest practical difference is the license: H3's license does not allow use in the EU, UK, US or South Korea, while LTX-2.5's does.
Side by side
| MiniMax H3 | LTX-2.5 | |
|---|---|---|
| Developer | MiniMax | Lightricks |
| Weights released | August 3, 2026 | August 2026 |
| Parameters | 33B | 22B |
| Clip length | 4–15 seconds | 6–20 s (API Fast at 720p/1080p); 6–10 s otherwise |
| Resolution, open weights | 768p native | Depends on hardware and pipeline |
| Resolution, hosted API | Up to 2K | Up to 4K |
| Frame rate | 24 fps | 24–50 fps (API) |
| Audio | Generated with the video, 32 kHz stereo | Generated with the video |
| Inputs | Text, first/last frame, up to 9 images, 3 videos and 3 audio clips as references | Text, image, video and audio; last frame on image and audio inputs (API) |
| Multi-shot in one generation | Shot cuts described in the prompt | Native multi-shot with consistent characters and voice |
| Aspect ratios | Seven on fal; six for H3 Max text to video | 16:9 and 9:16 (API) |
| ComfyUI | Native since August 3, 2026 | Native support |
| License | MiniMax H3 Community License | LTX-2.x Community License |
| Where the license applies | Worldwide except the EU, UK, US and South Korea | Worldwide, except sanctioned countries |
| Revenue threshold | Separate permission above $20M | Paid license above $10M annual revenue |
License: check this first
For many readers, the license decides it before quality does.
- MiniMax H3 grants rights only outside the European Union, the United Kingdom, the Republic of Korea and the United States. Running the weights there is not covered by the license. Companies with more than $20M in revenue need separate authorization, and Comfy notes that commercial use of locally generated outputs needs a MiniMax commercial license.
- LTX-2.5 has no such territory exclusion beyond sanctioned countries. Commercial and production use is free for organizations under $10M in annual revenue; above that, a paid license is required.
If you are in the US or Europe and want to run a model on your own GPU, LTX-2.5 is the one the license allows. Alibaba's Apache 2.0 licensed Wan 2.2 is another option — see MiniMax H3 vs Wan 2.2. Hosted APIs such as fal.ai come with their own terms of use, separate from the open-weights license.
Speed
LTX-2.5 is built for speed. Lightricks reports a 10-second image-to-video clip in 6.8 seconds on two NVIDIA GB200 chips. Community tests on the same consumer GPU report LTX-2.5 several times faster than H3 for a clip of the same length. Expect H3 to take minutes per clip on a single desktop GPU, even with Comfy's memory optimizations.
Quality and control
- H3 is the stronger choice when you need control over a specific shot: lock the first and last frame, or feed reference images, video and voice clips to keep a character consistent. Published hands-on comparisons generally rate H3 better on complex motion and physics.
- LTX-2.5 is the stronger choice for longer sequences: native multi-shot keeps characters, lighting and voice consistent across cuts, an optional duration predictor picks the clip length, and the API goes up to 4K.
No independent benchmark covers both models on the same prompts yet, so test your own use case on both.
Which one should you use?
| You want… | Pick |
|---|---|
| To run locally in the EU, UK, US or South Korea | LTX-2.5 |
| Fast iteration on a single GPU | LTX-2.5 |
| Clips longer than 15 seconds or 4K output | LTX-2.5 (API) |
| Multi-shot scenes from one prompt | LTX-2.5 |
| First and last frame control | MiniMax H3 |
| Character consistency from reference images, video and voice | MiniMax H3 |
| Wider choice of aspect ratios | MiniMax H3 |
Try H3 without a GPU
H3 Max, a faster post-trained version of H3, runs in the cloud on this site — no install, no GPU, and free credits for a first 480P video. Start with the H3 Max video generator, compare it with the base model in H3 Max vs H3, or set up the open weights with our MiniMax H3 ComfyUI guide. For API access and prices, see the H3 Max API overview.
FAQ
Is LTX-2.5 better than MiniMax H3? Neither wins outright. LTX-2.5 is faster, longer and higher-resolution through its API; H3 offers more control over individual shots and references.
Which model can I run locally in the US or EU? LTX-2.5. The MiniMax H3 Community License excludes the EU, UK, US and South Korea.
Do both models generate audio? Yes. Both generate sound together with the picture in one model, without a separate audio step.
Do both run in ComfyUI? Yes, both have native ComfyUI support.
Sources: MiniMax H3 model card and license, ComfyUI MiniMax H3 guide, LTX-2.5 model card, LTX-2.5 API docs, LTX-2.x Community License. Last updated: October 1, 2026.
h3maxvideo.net is independent and not affiliated with MiniMax, Lightricks or fal.ai.