MiniMax H3 Turbo: 4 vs 6 vs 8 Steps, and Which LoRA

Two weeks ago, running MiniMax H3 faster meant one experimental LoRA and a lot of hedging. Now there are four checkpoint lines, several ComfyUI conversions, quantized variants, and a steady stream of Reddit threads asking the same question: which one, and at how many steps?

This piece answers both. The short version is that the step count matters more than the checkpoint, and that the number in the name — 4 — is not the number most people should use.

What "Turbo" actually is

H3 Turbo is not an official MiniMax release. It is a set of community distillations that compress MiniMax H3's joint video-and-audio sampling from roughly 20 steps down to 4, and they are published as LoRAs you apply to the base model.

Two independent lines matter:

LightX2V / ModelTC publish under Apache-2.0 and cover FL2VA and T2VA at 544p, plus a 4-step v1.0 trained at 768p (1344×768) and a separate Ref2VA checkpoint. Their repository ships example ComfyUI workflows for text-to-video, image-to-video, and reference-to-video, which is the fastest way in if you do not want to wire a graph yourself.

larryvrh publishes a single-file LoRA line, currently at v4 step-600 EMA, around 744 MB in bf16. This is the line most of the Reddit discussion is about.

Everything else in circulation — Kijai's ComfyUI conversions, drbaph's pruned variants, the INT8 and CMF quantizations — derives from one of these two.

The step count is the real decision

Here is the thing that gets lost when a model is named after its step count: both authors recommend more than 4 steps.

The larryvrh model card puts 6–8 as noticeably better than 4, and reports that past 8 there is no further gain — over-sharpening appears instead of detail. ModelTC's own example workflows for T2VA and I2VA default to the 8-step checkpoint, not the 4-step one.

If your 4-step output looks smeared or ghosted on a moving subject, that is the documented failure mode, not a bad install. Raise the steps before you change checkpoints.

Which checkpoint

Reference-to-video has exactly one distilled option: the LightX2V Ref2VA 4-step checkpoint, still labelled v0.1. Treat its output as less settled than the FL2V line, and compare against a base run before shipping client work.

Everything else is a choice between two:

• larryvrh v4 step-600 EMA at 6–8 steps is the author's own recommendation and the better default. v4 fixed the over-sharpened, plastic look of the earlier ~850 line and clearly improved faces, fingers, and fine texture on static and small-motion shots.
• LightX2V 4-step v1.0 768p when you want maximum speed and native 1344×768 output rather than upscaling from 544p.

One genuine exception: at 4 steps with heavy, fast motion, the older larryvrh ckpt-850 still handles the motion better than v4 does. The author says so directly. If your work is action-heavy and you are pinned to 4 steps, that is the checkpoint to try.

Settings are the same across the board: LoRA strength 1.0, scheduler simple, and CFG left at 1.0 because these distillations are guidance-free.

About that "5× faster"

The roughly 5× figure is real, and it describes sampling steps: about 20 down to 4. It does not describe your wall clock.

Model loading, text encoding, VAE decode, and audio decode do not shrink alongside the sampler. A 5× cut in sampling produces something well short of a 5× shorter run, and how much short depends on how much of your time was sampling in the first place.

One user's numbers, on an RTX 5080 generating 15 seconds at roughly 0.5 MP:

That is a real speedup and worth having. It is also one person's machine and configuration, not a controlled benchmark — treat it as the shape of the gain rather than a number to expect.

The metric that actually decides whether Turbo is worth it is time per usable clip. A 4-step run you throw away because the motion smeared costs more than a 6-step run you keep. Counting only the runs that succeeded is how a 5× sampling cut turns into a 2× real speedup, or occasionally into none at all.

Where it still loses to base

Worth being straight about the limits:

• Fast motion at low steps smears and trails. This is the consistently reported weak spot.
• Audio quality is still being actively improved on the larryvrh line, per the author.
• Reference adherence on Ref2VA is the least settled part of the whole ecosystem, and it is still a v0.1 preview.

None of that makes Turbo a bad trade. It makes it a trade. For iteration, previews, and shot exploration, 4–8 steps changes how much you can try in an evening. For a final delivery with fast motion in frame, the base model is still the honest answer.

Getting started

Start from ModelTC's example workflows rather than building the graph yourself, pick the checkpoint that matches your task, and validate on a single clip before you queue a batch. If you want the checkpoint and step count picked for you based on task, priority, and VRAM, the MiniMax H3 Turbo selector does exactly that, with the trade-offs spelled out next to each recommendation.

If you do not have a GPU to run any of this on, the hosted MiniMax H3 generator on this site runs the standard model in the cloud — no Turbo LoRA, but no local setup either.

Checkpoint names, step recommendations, and settings in this article come from the LightX2V/ModelTC and larryvrh model cards and repositories, verified August 15, 2026. Turbo LoRAs are community releases under Apache-2.0, not official MiniMax products.