MiniMax H3 Highspeed vs H3 Lite: Which Model Should You Use?

Use MiniMax H3 Lite when the point is to explore cheaply. Use H3 Highspeed when you want MiniMax's middle 768P tier and are willing to pay more for the quality-and-speed balance MiniMax assigns to it. Standard H3 is still the only one of these direct API models with 2K output, while H3 Max has a different five-second minimum and its own prompt-expansion control.

That is the useful answer. The word “Highspeed” does not mean “cheapest,” and “Max” does not mean “highest resolution.” The current H3 menu makes more sense when you start with the shot's hard requirement instead of the model name.

H3 Highspeed vs H3 Lite—and where H3 and Max fit

MiniMax's live V2 schema lists four exact model IDs. Here is the current comparison as it stands on October 26, 2026:

These are direct MiniMax pay-as-you-go prices from the current official pricing table. They are not this site's retail credit values, and they do not include billable reference video or images beyond the free quota.

What an eight-second shot costs

An eight-second text-to-video job makes the difference easier to see. The same output length would cost:

For a plain text job, Lite at 768P is one-third cheaper than Highspeed: $0.32 instead of $0.48. Whether Highspeed is worth the extra $0.16 requires a controlled test on your own brief. The official pages do not publish a head-to-head benchmark for these variants, so treat their positioning as a reason to test—not as proof that every Highspeed clip will beat every Lite clip.

Pick the model from the constraint

If the final file must be 2K

Choose MiniMax-H3. Highspeed, Lite, and Max all stop at 768P in the current direct API. Paying for Max does not unlock 2K.

MiniMax also exposes a separate regeneration route that takes a 768P H3 result to 2K. Regenerated output is billed at $0.05 per second, and chargeable input materials from the original task are billed again. This is a second billable task, not a resolution switch on Lite, Highspeed, or Max.

If you are still finding the shot

Choose H3 Lite. A weak composition does not become a good idea because it has more pixels. Lite's 480P rate is suited to trying camera direction, timing, staging, and sound before paying for a more expensive pass.

A sensible ladder for one eight-second concept is:

1. Find the action at Lite 480P: $0.20 per attempt.
2. Check the selected direction at Lite 768P or Highspeed 768P: $0.32 or $0.48.
3. Move to standard H3 2K only if the selected deliverable needs it: $1.04.

This is a budgeting workflow, not a claim that a prompt transfers identically between models. Keep the prompt, duration, aspect ratio, and source media fixed while you compare, then judge the complete clip—including audio—not one flattering frame.

If four seconds matters

Use H3, Highspeed, or Lite. Their documented range starts at four seconds. H3 Max starts at five seconds, so a four-second request is outside its contract.

If you want the middle 768P option

Highspeed is the cleanest fit. MiniMax describes it as balancing generation quality and speed for everyday content creation. It sits between Lite 768P at $0.04 per second and H3 768P at $0.08 per second.

The name is a product tier, not a latency guarantee. Queue time, input transfer, encoding, and file delivery still affect the wait a user feels.

If you need Max's extra control

The direct MiniMax-H3-Max schema exposes prompt_expansion_mode with Balanced and Quality choices. It also uses a five-to-fifteen-second duration range. Those are concrete reasons to select Max. “It sounds like the biggest model” is not.

Reference jobs change the invoice

All four current direct models accept image, video, and audio references. That is easy to miss because Max is described as a multi-asset model, but references are not exclusive to Max.

There are two rules worth noticing:

• A first or last frame cannot be mixed with reference-role media in the same request. Pick an image-to-video job or a reference-to-video job.
• The flagship H3, Highspeed, and Lite tiers include the first five images in a task. Max includes the first two. Extra images and input video are billed at model-specific rates; audio input is currently free.

Suppose an eight-second 768P Max output uses a five-second reference video. The output is 8 × $0.08 = $0.64; the input video is 5 × $0.143 = $0.715. The job's published media charges total $1.355, before any extra images beyond the free quota.

The same arithmetic matters on every tier:

reference job cost =
  output duration × output rate
  + input-video duration × input-video rate
  + billable extra images × image rate

Do that calculation before choosing a model from its output price alone. A reference-heavy Max request can cost very differently from a text-only Max request of the same length.

Highspeed is not H3 Max Turbo

The naming collision matters if you are reading code or comparing invoices:

• MiniMax-H3-Highspeed is a model ID in MiniMax's direct V2 API.
• MiniMax-H3-Lite is another direct MiniMax model ID.
• minimax/h3-max-turbo/* names fal's post-trained H3 Max Turbo routes.
• FastH3 names a separate public-checkpoint family maintained through FastVideo.

AI Live Generator uses fal's H3 Max Turbo for its ordinary text and image jobs. It does not expose MiniMax's direct Highspeed or Lite IDs. The H3 Max versus H3 Max Turbo guide explains the fal routes, while the MiniMax H3 Max API guide covers the direct V2 request shape.

A practical routing card

Keep this beside the shot list:

• Cheapest 480P exploration: H3 Lite.
• Cheapest direct 768P tier: H3 Lite.
• MiniMax's middle 768P speed/quality tier: H3 Highspeed.
• Native 2K: standard H3.
• Four-second clip: H3, Highspeed, or Lite—not Max.
• Max prompt expansion or a Max-specific integration: H3 Max.
• Browser workflow on this site: fal H3 Max Turbo, with a different endpoint and credit schedule.

The best model is the least expensive one that satisfies the non-negotiable part of the brief. Start with resolution, duration, and reference inputs. The model name comes last.

MiniMax H3 Highspeed vs H3 Lite FAQ

Is H3 Highspeed the same as H3 Max Turbo?

No. Highspeed is a direct MiniMax model ID; H3 Max Turbo is a fal route and the ordinary model used by this site's generator.

Is H3 Lite an open-weight model?

The current MiniMax API documentation presents H3 Lite as a hosted pay-as-you-go model. Do not confuse it with FastH3's public checkpoints.

Which current H3 model is cheapest?

For published direct API output prices, H3 Lite at 480P is the lowest at $0.025 per second. At 768P, H3 Lite is $0.04 per second.

Does H3 Max produce the highest resolution?

No. Direct H3 Max tops out at 768P. Standard MiniMax-H3 is the current route with native 2K output.

Can H3 Lite and Highspeed use reference audio?

Yes. The current V2 schema lists reference image, video, and audio support for both. Audio input is listed as free, while input video and extra images can add charges.

The labels will probably keep multiplying. Exact model IDs, accepted inputs, and the pricing formula are the durable part of the decision.