MiniMax-H3 open weights and deployment

Status: Available. MiniMax published the H3-Base weights as two BF16 task checkpoints, FL2VA for text and first/last-frame input and Ref2VA for reference input; each task folder in the repository is about 144 GB. The release leaves out H3-Context-IR and H3-Regenerate-2K, so a self-hosted H3-Base generates 768p video and does not reproduce the MiniMax Platform's full 2K workflow. The MiniMax H3 Community License excludes the United States, European Union, United Kingdom and South Korea (organisations there can apply to MiniMax for a licence), requires separate written authorisation above US$20 million in yearly revenue and "MiniMax H3" displayed in commercial products, and forbids using outputs to improve other models.

Agent deployment prompt

You are my deployment agent. Deploy the official MiniMaxAI/MiniMax-H3 weights as a private video-generation service on this host. Use MiniMax's self-host guide and model card as the source of truth: https://platform.minimax.io/docs/guides/local-deploy-h3 and https://huggingface.co/MiniMaxAI/MiniMax-H3.

Licence
- The weights are under the MiniMax H3 Community License Agreement, which excludes the United States, the European Union, the United Kingdom and South Korea. Before downloading, show me its territory terms, its commercial terms (separate written authorisation above US$20 million in yearly revenue, and "MiniMax H3" displayed in commercial products) and its safeguard duties for hosted services, and ask me to confirm that this deployment and its users are outside the Excluded Territories or that MiniMax has granted us a separate licence. Stop if I cannot confirm.

Preflight
- Inspect the OS, GPU model and count, free VRAM, host RAM, driver and CUDA versions, free disk space and available ports. Each task checkpoint (FL2VA or Ref2VA) is about 144 GB in the repository. MiniMax publishes no minimum hardware: its guide pins one node with 8 × NVIDIA B200 as a reference configuration, not a minimum, and SGLang's upstream four-GPU H200 runs peaked at about 94 GB of VRAM per GPU. Estimate whether the official weights fit with safe headroom.
- If the hardware cannot safely run the official weights, stop before installing or launching anything. Explain the limiting resource and propose a separately reviewed alternative; do not substitute Comfy-Org's converted INT8/NVFP4 files or any other community quantization automatically.

Deployment
- If preflight passes and I have confirmed the licence, follow the guide's SGLang Quickstart in an isolated environment: install the SGLang source commit it pins, download only the FL2VA partition at the model revision it pins, and start the service with the repository root as --model-path and --model-variant fl2va. Reference-to-video needs a separate Ref2VA service; add it only if I ask.
- Bind to 127.0.0.1; do not expose the endpoint publicly. The licence requires authentication, quotas, audit logging and abuse controls before any hosted exposure, so list what that would take and ask me first. Keep any token or key you create in a protected environment variable.
- A self-hosted H3-Base produces 768p video. MiniMax's H3-Context-IR and 2K regeneration are not in the open release, so do not claim parity with the MiniMax API.

Verification and handoff
- Wait until /health returns {"status":"ok"}, then submit a 5-second text-to-video job to the asynchronous /v1/videos endpoint, poll it, download the MP4 and confirm with ffprobe one H.264 video stream at 24 FPS and one 32 kHz stereo AAC audio stream. Report the exact launch command, installed versions, model revision, local endpoint, test result, measured GPU memory and log location. Identify any reverse-proxy or firewall changes that still need my approval. Never print tokens or other secrets.
License
MiniMax H3 Community License Agreement (excludes the US, EU, UK and South Korea)
Hardware note
Not published as a minimum. MiniMax's self-host guide pins one node with 8 × NVIDIA B200 as a reference configuration and says it is not a minimum; the model card's example uses four GPUs without naming them. SGLang's upstream benchmark generated a 5-second 1344×768 clip on 4 × H200 in about 75 seconds at about 94 GB peak VRAM per GPU; its consumer-GPU offload profiles are not MiniMax-verified.
Recommended runtime
SGLang (MiniMax's pinned self-host baseline); vLLM-Omni and ComfyUI are also documented, and ComfyUI uses Comfy-Org's converted INT8/NVFP4 files rather than the official checkpoint
Verified date
2026-09-25

Official repository

Published model card

Parameter count
33B-parameter dense Omni-Transformer, about 13B of it in AdaLN branches that an inference-only deployment does not need to load
Context window
Not applicable (video model); prompts are limited to 7,000 characters
Maximum output
4–15 seconds at 24 FPS with 32 kHz stereo audio; 768P or 2K through the API, 768p from the open H3-Base weights
Input/output modalities
Text with an optional first frame, last frame or both, or with up to 9 reference images, 3 reference video clips and 3 reference audio clips (frame and reference inputs cannot be mixed); video with synchronized stereo audio output

Official sources