~744 MB
Turbo adapter
bf16 LoRA weights
MiniMax H3 Turbo LoRA
Free few-step audio-video generation in your browser
MiniMax H3 Turbo LoRA accelerates the open MiniMax H3 base model while preserving its joint video and synchronized stereo-audio workflow. Use the embedded public demo for text-to-video, first-frame or first-and-last-frame experiments without installing the 33B base model.
Hugging Face ZeroGPU 4–8 sampling steps Text or keyframes Synchronized stereo audio
Standard H3 about 20 steps
Turbo LoRA 4–8 useful steps
Up to about 5× fewer sampling steps; wall-clock speed still depends on queue, canvas and hardware.
Try MiniMax H3 Turbo LoRA online
The official MiniMaxAI-hosted Space is embedded below. It currently exposes prompt, optional first and last frames, canvas, duration, steps, seed, prompt upsampling and LoRA selection.
Free public demo, not unlimited compute
Generation runs on Hugging Face ZeroGPU. Queues, daily user quotas, cold starts and temporary downtime are expected, and the Space may request Hugging Face sign-in for more capacity.
MiniMaxAI Space · ZeroGPUminimaxai-minimax-h3-turbo-lora.hf.space
Loading the MiniMax H3 Turbo LoRA demo…
What is MiniMax H3 Turbo LoRA?
It is a low-rank adapter for MiniMax H3, not a separate foundation model. The adapter changes the denoising behavior so H3 can produce useful results with far fewer sampling steps while retaining joint picture-and-sound generation.
Adapter, base model and demo are three different things
The roughly 744 MB LoRA still needs the much larger MiniMax H3 base, VAEs and text encoder for local use. The free Space packages that stack on remote GPUs; installing only the LoRA file cannot generate video by itself.
~33B
H3 base model
still required locally
24 fps
Native frame rate
current implementation
5–15 s
Validated duration
frame-grid aligned
Few-step sampling
Four steps are the fast design point; six to eight generally trade more compute for cleaner detail and motion.
Video and audio together
H3 denoises visual and audio streams together, producing a synchronized stereo soundtrack rather than a separate dubbing pass.
Text and keyframe control
The public workflow accepts text alone, a first image, or first and last images for guided image-to-video generation.
Open local workflow
The adapter can be used with its Apache-2.0 weights and maintained ComfyUI custom nodes, but the large base model still drives hardware needs.
Recommended MiniMax H3 Turbo LoRA settings
These recommendations come from the current Turbo LoRA model card. They are most useful for local ComfyUI runs; controls exposed by the public Space may change.
| Control | Recommended | Why |
|---|---|---|
| Checkpoint | v4 step-600 EMA | Best general checkpoint for static shots, small motion and fine detail. |
| Sampling steps | 6–8 | Usually cleaner than four steps; going beyond eight offers little benefit and can oversharpen. |
| LoRA strength | 1.0 | The adapter is tuned around 1.0. Change it only to correct a specific artifact. |
| Scheduler | simple | The published workflow and sampler are designed around the simple scheduler. |
| Fast draft | 4 steps | Useful for iteration, but fast or intense motion can show smear or trailing artifacts. |
Free Space API and production API
The embedded Gradio Space is callable as an API, but it remains a public demo without uptime guarantees. A paid provider is more appropriate for production traffic.
Public Gradio API
Inspect the live schema with view_api() before sending a job. The current workflow endpoint and generic parameter names can change when the Space UI is updated.
Current endpoint: /output_videoSchema checked 4 September 2026
import { Client } from "@gradio/client";
const client = await Client.connect(
"MiniMaxAI/MiniMax-H3-Turbo-Lora"
);
// Read the live schema before calling the public Space.
console.log(await client.view_api());Paid hosted fallback
WaveSpeed currently offers a dedicated MiniMax H3 text-to-video LoRA REST endpoint with 3–15 second output, custom LoRAs and native stereo audio.
480p · $0.05 / generated second768p · $0.10 / generated second
Price snapshot checked 4 September 2026. Verify the provider before a paid run.
What to know before using it
Turbo reduces sampling work; it does not remove the operational limits of a large audio-video model.
ZeroGPU is shared
Anonymous capacity is small and queues can be long. A successful run is never guaranteed.
The base model remains large
A small LoRA download does not make local H3 lightweight; the base model, encoders and VAEs still dominate storage and memory.
Fast motion needs care
Four-step generation can smear fast subjects. Increase steps and simplify the scene before changing many other controls.
Public API is not an SLA
Endpoints, inputs, availability and authentication behavior can change. Do not build a critical backend around a community quota.
MiniMax H3 Turbo LoRA FAQ
Is MiniMax H3 Turbo LoRA a new video model?+
No. It is a LoRA adapter applied to the MiniMax H3 base model to reduce the useful sampling-step count.
Can I use the online generator for free?+
The embedded public Space uses Hugging Face ZeroGPU and can be tried without a MiniMax API key. User quotas, queues and sign-in prompts may apply.
Does it support image-to-video?+
Yes. The current workflow accepts an optional first frame and optional last frame in addition to a text prompt.
Does it generate audio?+
Yes. MiniMax H3 generates synchronized stereo audio together with the visual stream; audio quality is still described as an area under improvement.
What settings should I start with?+
For the current v4 checkpoint, start with six steps, strength 1.0 and the simple scheduler. Use eight steps when quality matters more than speed.
Can I use the free Space as my production API?+
It is technically callable, but it has public quotas and no service guarantee. Use a paid provider or self-host for production workloads.
Continue exploring audio-video tools
Technical sourcesVerified 4 September 2026 MiniMaxAI Space Turbo LoRA model card Hugging Face Space API Hugging Face embedding WaveSpeed LoRA API