---
title: MiniMax-H3-NF4
canonical_url: "https://www.modelscope.cn/studios/MiniMax/MiniMax-H3-NF4"
md_url: "https://www.modelscope.cn/studios/MiniMax/MiniMax-H3-NF4.md"
repository: MiniMax/MiniMax-H3-NF4
chinese_name: MiniMax-H3-NF4
last_updated: 2026-09-02
sdk_type: gradio
sdk_version: 6.17.3
downloads: 0
stars: 99
---

# MiniMax-H3-NF4

> MiniMax-H3-NF4 - MiniMax 在 ModelScope 创建的在线 Demo。基于 MiniMax-H3-NF4 与 DiffSynth 的视频和同步音频生成体验，支持 Larry Turbo v4、LightX2V Turbo v0.1 两个 LoRA，在 ModelScope xGPU 上运行。

MiniMax/MiniMax-H3-NF4 是 ModelScope 魔搭社区上的在线可交互 Demo（创空间），基于 gradio 6.17.3 构建。

- **Repository**: MiniMax/MiniMax-H3-NF4
- **SDK**: gradio
- **SDK version**: 6.17.3
- **Downloads**: 0
- **Stars**: 99
- **Last updated**: 2026-09-02

Source: https://www.modelscope.cn/studios/MiniMax/MiniMax-H3-NF4

---

# MiniMax-H3 NF4 Turbo — ModelScope xGPU

基于 `DiffSynth-Studio/MiniMax-H3-NF4` 的视频与同步音频生成创空间，
运行于 ModelScope xGPU。支持 Larry Turbo v4 与 LightX2V Turbo v0.1 两个
LoRA，可按场景切换快速推理，也可关闭 LoRA 使用 NF4 基模。

魔搭 AIGC 专区现已支持 MiniMax H3 的训练与推理。如需体验完整的视频生成能力与更精细的参数控制，可前往
[H3 视频生成](https://modelscope.cn/aigc/video-generation) 或
[模型训练](https://modelscope.cn/aigc/model-training)。

## Backend

- Base model: [`DiffSynth-Studio/MiniMax-H3-NF4`](https://modelscope.cn/models/DiffSynth-Studio/MiniMax-H3-NF4)
- Runtime: ModelScope xGPU, CUDA/bfloat16 compute with NF4 weights
- Memory: DiffSynth disk → CPU → GPU offload, with model cache under `/mnt/workspace`
- Conditioning: local NF4 MiniMax-H3 text encoder
- Turbo LoRA: Larry, LightX, or Off, loaded lazily from ModelScope. Both LoRAs
  are applied as runtime residuals and are never merged into the quantized weights.

## Compatibility note

The UI provides keyframe cover-crop, canvas/duration/steps/seed controls,
synchronized MP4 output, and LoRA switching. Prompt upsampling is currently a
no-op; inference uses the prompt entered by the user.

## Environment variables

| Name | Default | Purpose |
|---|---|---|
| `H3_MODEL_REPO` | `DiffSynth-Studio/MiniMax-H3-NF4` | NF4 checkpoint |
| `H3_PROCESSOR_REPO` | `MiniMax/MiniMax-H3` | FL2VA processor |
| `H3_WORKSPACE` | `/mnt/workspace` | Persistent cache/output root |
| `H3_MAX_OUTPUTS` | `5` | Number of newest generated MP4 files retained; older videos are deleted automatically |
| `H3_GPU_RESERVE_GB` | `4` | GPU memory kept free from model placement |
| `H3_LORA_REPO` | `larryvrh/MiniMax-H3-Turbo-Lora` | Larry LoRA model |
| `H3_LORA` | `minimax_h3_turbo_v4_step600_ema.safetensors` | Larry v4 step600 EMA LoRA file |
| `H3_LIGHTX_REPO` | `lightx2v/Minimax-h3-Turbo` | LightX LoRA model |
| `H3_LIGHTX_FILE` | `minimax_h3_fl2v_turbo_4step_v0.1.safetensors` | LightX LoRA file |

## Experimental LoRA validation

Before a LoRA is attached, the runtime now requires complete A/B pairs, 100%
module-name coverage, and compatible matrix shapes. Larry defaults to 6 steps,
LightX defaults to 4, and the unmodified NF4 base defaults to 50 when the UI
selection changes. API callers may still explicitly choose a different step
count. LightX uses a local Diffusers/PEFT-to-DiffSynth conversion with Q/K/V
triplets fused into DiffSynth's `qkv_proj`; it remains experimental until its
output has been compared against the official BF16 Diffusers implementation.

Before and after every request (including failed requests), all DiffSynth
VRAM-managed models are offloaded using the pipeline's normal device manager,
then Python and CUDA caches are collected. Memory statistics are emitted as
`[vram]` log lines so repeated-request behavior can be audited.

Generated videos are stored under `/mnt/workspace/h3-outputs`. At startup and
before each video write, the app removes the oldest outputs beyond
`H3_MAX_OUTPUTS` (five by default). Incomplete MP4 files are removed when video
encoding fails. Model files under `/mnt/workspace/modelscope-cache` are never
included in this cleanup.