Deploy MiniMax H3

MiniMax H3 is a native joint video-and-audio omni-modal model. The open H3-Base release supports text-to-audiovisual generation, first/last-frame control, and image, video, and audio reference conditions, with synchronized outputs up to 15 seconds.

H3-Context-IR is not part of the open release. A local H3-Base deployment can generate 768p output, but it cannot independently reproduce the complete MiniMax Platform H3-Context-IR and 2K Workflow.

Open release and system boundary

The complete H3 system contains three parts:

Component	Included in the open release	Purpose
H3-Context-IR	No	Interprets free-form text, image, video, and audio context and produces a structured representation for H3-Base
H3-Base FL2VA / Ref2VA	Yes	Generates 768p video with synchronized stereo audio
H3-Regenerate-2K	No; the complete local workflow also uses APIs	Regenerates 2K video from the original context and 768p result

Resource	Description
Model repository	`MiniMaxAI/MiniMax-H3`
ModelScope	`MiniMax/MiniMax-H3`
License	MiniMax H3 Community License Agreement
SGLang Cookbook	MiniMax-H3

MiniMax H3 uses a dedicated Community License that includes territory, commercial-use, distribution, attribution, and Acceptable Use Policy terms. The current license lists the United States, European Union, United Kingdom, and South Korea as Excluded Territories and provides a related application process. Read the complete license and determine its applicability before downloading, deploying, or using the model commercially.

Choose a checkpoint

H3-Base contains two separate checkpoint partitions. Run separate FL2VA and Ref2VA servers to provide every capability.

Checkpoint	`--model-variant`	Request tasks	Input conditions
FL2VA	`fl2va`	`t2va`, `fl2va`	Text, with an optional first frame, last frame, or both
Ref2VA	`ref2va`	`ref2va`	Text with image, video, or audio references; video-to-video is part of this task

Do not point --model-path to a manually downloaded FL2VA/ or Ref2VA/ subdirectory. Use the root model ID and select the partition with --model-variant.

Output specification

Property	Specification
Duration	4–15 seconds
Resolution	768-pixel short edge
Frame rate	24 FPS
Audio	32 kHz stereo
Container	MP4 with H.264 video and AAC audio

Choose hardware

The SGLang Cookbook provides an interactive generator and continuously maintains more hardware, quantization, offload, and parallelism combinations. Documented reference configurations include:

Hardware	Reference approach	Notes
8 × B200	Ulysses 8, resident weights	Lossless reference configuration used on this page
4 × H200	Ulysses 4, resident weights	Low-latency configuration measured by SGLang
4 × H100 80 GB	TP 2 + Ulysses 2	Balances memory and speed
2 × RTX 5090 32 GB	TP 2 + layerwise offload	Requires approximately 384 GiB-class host memory
1 × RTX 4090 24 GB	INT8 + layerwise offload	Capacity path; slower and not a lossless reference

Use the MiniMax-H3 configurator to generate the complete command for your target hardware. The generator marks combinations without completed end-to-end validation as Unverified.

Reference deployment: 8 × B200, FL2VA

Install SGLang Diffusion

pip install -U uv
uv venv --python 3.12
source .venv/bin/activate
uv pip install "sglang[diffusion]" --prerelease=allow

Start the server

sglang serve \
  --model-path MiniMaxAI/MiniMax-H3 \
  --model-variant fl2va \
  --num-gpus 8 \
  --ulysses-degree 8 \
  --encoder-parallel auto \
  --performance-mode speed \
  --host 0.0.0.0 \
  --port 30010

To download from ModelScope, set the environment variable and replace the model ID:

SGLANG_USE_MODELSCOPE=true sglang serve \
  --model-path MiniMax/MiniMax-H3 \
  --model-variant fl2va \
  --num-gpus 8 \
  --ulysses-degree 8 \
  --encoder-parallel auto \
  --performance-mode speed \
  --host 0.0.0.0 \
  --port 30010

Start Ref2VA as a separate service on a different port:

sglang serve \
  --model-path MiniMaxAI/MiniMax-H3 \
  --model-variant ref2va \
  --num-gpus 8 \
  --ulysses-degree 8 \
  --encoder-parallel auto \
  --performance-mode speed \
  --host 0.0.0.0 \
  --port 30011

Generate text-to-video-and-audio

H3 uses an asynchronous OpenAI-compatible video endpoint. Create a task first:

curl -X POST http://127.0.0.1:30010/v1/videos \
  -H "Content-Type: application/json" \
  -d '{
    "model": "MiniMaxAI/MiniMax-H3",
    "prompt": "At night, three cats perform with tiny brass instruments in a warm living room, with movement synchronized to the music.",
    "seconds": 5,
    "task": "t2va",
    "conditions": [],
    "target": {
      "short_edge": 768,
      "aspect_ratio": "16:9",
      "duration_seconds": 5.0
    },
    "num_outputs_per_prompt": 1,
    "num_inference_steps": 50,
    "flow_shift": 12.0,
    "audio_flow_shift": 3.0,
    "seed": 1101
  }'

The response returns a task ID. Use it to query status and download the output:

curl http://127.0.0.1:30010/v1/videos/<video_id>

curl -L http://127.0.0.1:30010/v1/videos/<video_id>/content \
  -o minimax-h3-t2va.mp4

Download the result after the task status becomes completed. For a failed task, inspect the service logs for GPU memory, input, and checkpoint-partition errors.

Keyframes and reference media

First- or last-frame control: call the FL2VA server with task="fl2va" and one or two image conditions with role="keyframe".
Image, video, or audio references: call the Ref2VA server with task="ref2va".
Video-to-video is a Ref2VA input form and does not use a separate task value.
A Ref2VA request must contain at least one reference condition. Use FL2VA for keyframes alone.

Reference media can use server-local files. Mount the media directory read-only and restrict accessible paths to prevent arbitrary file access.

Quality and performance

The B200 configuration on this page uses the official mixed BF16/FP32 weights and a lossless reference path.
Quantization, approximate attention, Cache-DiT, fewer inference steps, and some community weights can change output quality.
Do not enable an unverified approximate optimization when generating consistency baselines.
The initial H3 open release and the complete platform workflow have different component boundaries. Record the checkpoint, SGLang version, all sampling parameters, and seed when comparing results.

Resources

SGLang MiniMax-H3 Cookbook

Open the complete hardware generator, request modes, optimizations, and benchmarks.

MiniMax-H3 model repository

Open the model card, checkpoints, workflows, and license.

Get started

Use Guides

FAQs

Terms & Policy

Open release and system boundary

Choose a checkpoint

Output specification

Choose hardware

Reference deployment: 8 × B200, FL2VA

Install SGLang Diffusion

Start the server

Generate text-to-video-and-audio

Keyframes and reference media

Quality and performance

Resources

SGLang MiniMax-H3 Cookbook

MiniMax-H3 model repository

​Open release and system boundary

​Choose a checkpoint

​Output specification

​Choose hardware

​Reference deployment: 8 × B200, FL2VA

​Install SGLang Diffusion

​Start the server

​Generate text-to-video-and-audio

​Keyframes and reference media

​Quality and performance

​Resources

SGLang MiniMax-H3 Cookbook

MiniMax-H3 model repository

Open release and system boundary

Choose a checkpoint

Output specification

Choose hardware

Reference deployment: 8 × B200, FL2VA

Install SGLang Diffusion

Start the server

Generate text-to-video-and-audio

Keyframes and reference media

Quality and performance

Resources