MiniMax H3 Private Deployment Guide: Licensing, Hardware, and Inference Recipes

Forum topic · ✨步子哥 · 2026-09-07

Summary

This in-depth research note covers private (on-premise) deployment of MiniMax H3, an open-source video generation model with native stereo audio released on 2026-07-31. Only the H3-Base 768p backbone is open-sourced (FL2VA and Ref2VA checkpoints); the 2K upscaler (H3-Regenerate-2K) and prompt-understanding module (H3-Context-IR) remain API-only. The community license excludes the EU, UK, Korea, and the US but covers mainland China, with commercial use free below $20M annual revenue (above that, written authorization is required) and mandatory 'MiniMax H3' UI attribution. The model comprises a 33B dense transformer with a ~13B AdaLN branch, a Qwen3-VL-32B encoder (Apache 2.0), and dual VAEs — roughly 134 GiB BF16 per FL2VA partition. Deployment paths: SGLang (production choice; 4×H100/H200, 8×Ascend NPU recipes verified), vLLM-Omni (2×RTX 5090 verified, OpenAI-compatible API), and ComfyUI for consumer GPUs (8–24GB). Key engineering notes: H3 is compute-bound so batching gives no benefit, CFG-distilled weights require cfg-parallel-size=1, and Cache-DiT/Turbo LoRA deliver 1.3–6× speedups. A hybrid local-768p + API-2K workflow is the only route to official 2K quality.

Key points

What H3 is: MiniMax H3 (open-sourced 2026-07-31) is a general omni-modal *generation* system producing 4–15s, 24 FPS video with native 32kHz stereo audio (768px short edge default, 11 prompt languages). Two open checkpoints: FL2VA (text-to-video + first/last frame control) and Ref2VA (up to 9 images + 3 videos + 3 audios as references). Only H3-Base (768p backbone) is open; H3-Context-IR (prompt understanding) and H3-Regenerate-2K (768p→2K) are API-only.
Architecture: H3-Omni-Transformer is a 33B dense single-stream model (50 blocks) with ~13B params in an AdaLN branch (precomputable/skippable at inference; community "pruned" quantizations cut ~40% size). Encoder is Qwen3-VL-32B (Apache 2.0, ~46–51.5 GB BF16). Dual VAEs: VisualVAE (32× spatial / 4× temporal compression) + AudioVAE. A full FL2VA partition is ~134 GiB BF16; plan ≥500GB NVMe storage.
License: MiniMax H3 Community License (v2026-08-02, grantor Nanonoble Pte. Ltd., Hong Kong law). Applies globally except EU, UK, Korea, US; mainland China is covered. Commercial products with >$20M annual revenue need written authorization (api@minimax.io). Mandatory prominent "MiniMax H3" UI attribution (Section IV.2). Cannot use H3 or its outputs to train other models; no military use; self-hosted services must provide safety guardrails (plus China's AIGC/deep-synthesis compliance obligations). HuggingFace repo is currently ungated; ModelScope mirror MiniMax/MiniMax-H3 available (SGLang supports SGLANG_USE_MODELSCOPE=true).
Route A — SGLang (production choice): sglang serve --model-path MiniMaxAI/MiniMax-H3 --num-gpus 4 --ulysses-degree 4 --performance-mode speed --model-variant fl2va
4×H100: TP2+Ulysses2, peak ~57GB/card (fastest lossless recipe)
4×H200: 1344×768/124 frames/50 steps lossless median 85.5s; quality=high (Cache-DiT) 63.4s (1.35×, SSIM 0.9709)
8×H200: up to 6.24× speedup with quality loss (SSIM 0.76–0.91)
Single RTX 4090 24GB: kitchen_int8, peak ~18GB, 1011×576/4s in 130.7s
Huawei Ascend NPU verified: 8-card TP2+SP4 + Laser Attention + Cache-DiT; 8-card is 46.8% lower E2E latency than 4-card — a validated domestic-hardware path
Route B — vLLM-Omni (v0.26.0+, OpenAI-compatible /v1/videos): 2×RTX 5090 (32GB) with TP2 + layer-wise offload: 1344×768/124-frame/50-step in 8m38s, peak 22.6 GiB/card. Online FP8 cuts peak memory 68.52→53.51 GiB (-22%). Key ops finding: H3 is compute-bound, not bandwidth-bound — request batching gives no benefit and raises tail latency; keep --max-num-seqs 1. Ref2VA support is narrower than SGLang's.
Route C — ComfyUI / consumer: 24GB+ → pruned AdaLN INT8 DiT (19.53 GiB) + TE nvfp4_awq (14.61 GiB); 12–16GB → GGUF Q4_K_M (10.64 GiB) + TE Q2_K; 8GB → DiffSynth NF4 (slow). Apple Silicon: antirez/h3.c (MIT, Metal-native). Turing 2080Ti 22G has a W4A8+Turbo path. Turbo LoRA: 50 steps → 6–8 steps sweet spot (4 steps degrades audio/fast motion). FastH3 (DMD2): 6.9× denoising on 8×B300.
Route D — Hybrid (only way to official 2K): generate 768p locally, call official Context-IR (structured prompts) and Regenerate-2K APIs with prompts/drafts only. 2K is true regeneration from context, not classic upscaling. Pricing: 768p $0.08/s + 2K regen $0.05/s = $0.13/s, same as direct 2K.
Cost math: 4×H200 lossless ≈ $17.4/hour equivalent output, roughly break-even with cloud rental; quality=high or Turbo/FastH3 makes self-hosting 1.5–4× cheaper. Self-hosting pays off for high volume, data sovereignty, fine-tuning, or avoiding moderation constraints.

Pitfalls

2K and Context-IR are not open-sourced; no 2K self-hosting path.
No sparse-attention inference yet (training used it); long sequences pay in VRAM/speed.
CFG-distilled weights: cfg-parallel-size > 1 is rejected.
Batching is counterproductive (unlike LLM serving intuition).
vLLM-Omni known issues: U2×R2 mask bug (use pure Ulysses), VAE tile-mode only, incomplete Ref2VA reference support.
Turbo LoRAs are filename-validated; 4-step tier degrades audio.
License is accepted on download; keep UI attribution current and watch for territory changes.

Sources

Open-source announcement | HF model card | License | License Q&A | Self-hosting guide | awesome-minimax-h3-integration | ModelScope mirror
SGLang cookbook | vLLM-Omni recipe | LMSYS 8×H200 blog | ComfyUI tutorial | NVIDIA Sol-Engine H3 Super Acceleration | Reuters coverage