H3 Lite is UlazAI's name for a distilled, post-trained build of MiniMax H3 available through its Video Studio. Its main appeal is a lower per-second credit rate, with narrower reference support than UlazAI's full H3 offering. The public product description does not identify a downloadable checkpoint or establish a new MiniMax official model release.
This guide checks the H3 Lite product and example page, credit packages, and full H3 comparison page on October 16, 2026. All test outcomes discussed here belong to the provider. H3Vid has not run H3 Lite or integrated this named offering.
What the lower-priced option changes
UlazAI lists text-to-video, image-to-video, first-and-last-frame generation, and a limited reference route for five-to-fifteen-second clips with sound. Its page describes 2K as upscaled output. A delivery file labeled 2K therefore should not be read as proof of native 2K diffusion or a measured detail advantage.
The named Lite option is also separate from generic Turbo adapters, FastH3 packages, and the LightVAE decoder. Those names refer to different components or services. A shared goal of lower cost or faster processing does not establish shared weights, equivalent inputs, or interchangeable workflows.
If you want to reproduce a provider result elsewhere, first identify the actual execution model, acceleration recipe, prompt rewriting, and upscale path. The public Lite page provides a usable product description, but does not expose all of that provenance. Do not turn its output examples into instructions to download an assumed equivalent checkpoint.
Calculate credits for the selected resolution
The checked list rate is six UlazAI credits per output second at 768p, 1080p, or 2K. At 4K, the rate becomes thirty. The difference is fivefold, even if the creative brief stays the same.
| Requested length | Up to 2K | 4K |
|---|---|---|
| 5 seconds | 30 credits | 150 credits |
| 10 seconds | 60 credits | 300 credits |
| 15 seconds | 90 credits | 450 credits |
These are arithmetic examples from provider credit rates, not dollar quotes or observed invoices. Credit-package purchases determine how much those credits cost an account. Promotional balances and package bonuses can change an effective dollar comparison, so compare both services on the same basis before declaring a universal savings percentage.
UlazAI's full H3 page lists 36.5 credits per generated second and separate reference-video charges. That is the comparison behind the Lite page's approximate sixfold output-rate difference. It is not a general claim that every H3 provider charges the same amount, or that Lite preserves every feature of full H3.
Choose one visual conditioning route
For image-to-video, one photo supplies the start, with an optional second image for the ending. For reference-to-video, the page permits up to four reference images or one reference video. It does not offer audio-reference input. Generated dialogue and ambient sound are different from uploading an approved recording to be preserved.
Start by deciding what the source is meant to control. A first image defines an opening composition; semantic reference pictures describe appearance; a driving clip may dominate movement and scene content. These roles should be explicit in the prompt instead of being inferred from file order alone.
The provider reports a failed experiment mixing a robot image with a dancer video: the dancer remained instead of becoming the robot. That single result does not prove mixed references never work, but it is a reason to begin with one reference category and inspect identity carefully. If you need the full model's broader multimodal inputs, compare that contract directly rather than assuming Lite accepts the same request.
Understand the prompt planner
The example page says UlazAI turns its planner on for prompts shorter than three hundred characters and off for detailed shot lists. It reports that the same short cat-DJ prompt produced different subjects with planner on and off. That is evidence about a whole service pipeline, including rewriting, rather than a pure checkpoint comparison.
For a controllable first experiment, write one subject, one action, one setting, and a sound instruction. Save the exact prompt and planner setting. If a run misses the subject, change either the wording or planner state first, keeping other choices fixed. An original brief might ask for a small ceramic bird rotating once on a wooden shelf, with fixed framing and quiet room tone. This is a suggested test design, not a demonstrated Lite result.
API access now has named Lite routes
Update, October 23: UlazAI's Video Generation API documentation now lists h3_lite_t2v, h3_lite_i2v, and h3_lite_r2v. The published entries include five-to-fifteen-second choices, 768p/1080p/2K/4K quality modes, and the Lite credit schedule. This resolves the earlier question of whether a named Lite API offering was documented; it does not identify the underlying downloadable checkpoint.
The shared API requires a bearer key and a registered model_slug. It directs integrations to fetch the current video-model capabilities before submitting, then poll the Video Studio status route with the returned job ID. Use the selected model's current input contract rather than applying a different model family's reference-array fields. The generic API examples cover several families, so they are not proof that every Lite mode supports the same reference combination.
This guide has not called those generation routes or established a dollar invoice, refund policy, or upstream weight identity for Lite. H3Vid still does not provide this offering; the new information is documented provider access, not a completed integration here.
Read the examples as selected provider evidence
UlazAI describes a thirty-three-job run on October 14, with eight jobs submitted concurrently. Its timing includes queueing through completed download. That context matters when comparing a multi-minute turnaround with another service's model-only sampling number.
The page also publishes misses involving identity, extra subjects, physical causality, and unwanted text. Successful rendering is therefore different from satisfying the prompt. A completed file can still need another take or editorial repair. Review the full clip and sound before judging whether the cheaper output rate produced a cheaper usable shot.
To try the service, follow its Lite studio link, which currently requires sign-in. Begin with a short chosen mode and inspect its account-specific price before submitting. For H3Vid's separate online reference workflow, see MiniMax H3 Ref2VA. Neither that page nor this explanation implies that H3 Lite is already available in H3Vid's generator.