---
title: MiniMax-H3 Turbo-SLA
emoji: 🎬
colorFrom: pink
colorTo: blue
sdk: gradio
sdk_version: 6.25.0
app_file: app.py
pinned: false
short_description: 4-step distilled image-to-video with SLA sparse attention
python_version: "3.12"
startup_duration_timeout: 1h
models:
  - MiniMaxAI/MiniMax-H3
  - lightx2v/Minimax-h3-Turbo-SLA
---

# MiniMax-H3 Turbo-SLA

4-step distilled FL2V (first-and-last-frame to video) generation with **SLA (Sparse-Linear Attention)**
for efficient inference, producing video with a synchronized soundtrack.

This demo applies the [lightx2v/Minimax-h3-Turbo-SLA](https://huggingface.co/lightx2v/Minimax-h3-Turbo-SLA)
LoRA — a 4-step turbo distillation checkpoint with 85% attention sparsity — on top of the
[MiniMaxAI/MiniMax-H3](https://huggingface.co/MiniMaxAI/MiniMax-H3) base model.

## Architecture

This Space is the **denoising half** of a split MiniMax-H3 deployment:

- The 62 GiB Qwen3-VL text encoder/conditioner runs in
  [`multimodalart/qwen3vl-conditioner`](https://huggingface.co/spaces/multimodalart/qwen3vl-conditioner),
  called over the gradio API for each request.
- This Space loads the 61.7 GiB transformer + 10.4 GiB VAEs (bfloat16, unquantized)
  and applies the SLA Turbo LoRA at startup.

The split is necessary because MiniMax-H3 is 195.9 GiB in bfloat16 and a ZeroGPU Space
is evicted at 150 GB of storage.

## SLA (Sparse-Linear Attention)

The SLA LoRA was trained with [SLA](https://github.com/thu-ml/SLA) — a fine-tunable
sparse-linear attention mechanism that achieves approximately 2.5× inference
acceleration on RTX 5090 with 85% attention sparsity ratio while retaining
competitive visual quality.

> **Note:** The original LightX2V inference path uses SageAttention ("sage2") for the
> sparse attention kernel. ZeroGPU's Blackwell (sm_120) hardware does not support
> FA3/SageAttention kernels (sm90-only), so this Space uses cuDNN's fused attention
> backend instead. The LoRA weights are applied identically; only the attention
> kernel differs.

## Usage

Provide a text prompt and optionally a first frame and/or last frame image to
generate a 5-second video with synchronized stereo audio. The 4-step turbo
distillation makes each generation take about 30-60 seconds.