Deploy MiniMax H3 on GPU Cloud: Open-Weight Video Model Setup (2026)

MiniMax unveiled H3 at WAIC 2026 on July 17 (source), then shipped it through the API on July 31, 2026, as "a general-purpose multimodal generation model that reads text, images, video, and audio as one unified context and returns video with native stereo sound" (source). Days later, open weights followed under a license that flatly excludes the US, EU, UK, and South Korea from local deployment. Deploying MiniMax H3 on GPU cloud means understanding two things before you download a single checkpoint: the open weights only generate base 768p video, not the native 2K output the API demos show off, and if your organization operates in one of those four regions, self-hosting it isn't licensed without individual authorization. This guide covers what actually ships open, real VRAM numbers from consumer to datacenter GPUs, the license restriction in plain terms, and a working ComfyUI deployment on Spheron. H3 is a separate bet from MiniMax's text side; if you're deploying both, our MiniMax M3 guide covers the 1M-context coding model that shipped two months earlier.

What Is MiniMax H3 (H3-Omni Transformer, 2K Native-Audio Video Generation)

The H3-Omni Transformer at the center of MiniMax H3 is a dense, single-stream model: 33 billion parameters across 50 layers with a hidden size of 5,376 and 56 attention heads. It uses Qwen3-VL-32B as its text encoder and 3D multimodal RoPE across temporal, height, and width dimensions (source). Roughly 13B of those 33B parameters sit in AdaLN branches, which matters once you get to the pruned checkpoints below.

Output specs: 4-15 second clips in integer durations only, 24fps, native 32kHz stereo audio generated in the same pass as the video (not bolted on afterward), aspect ratios from 21:9 to 9:16, and support for 11 languages (source). A single generation can take up to 9 reference images, 3 video clips (2-15 seconds each), and 3 audio clips, 12 files total, with prompts capped at 7,000 characters (source).

Four architectural pieces do the actual work:

• Contextual Omni Representation distills roughly 100K tokens of source material down to an average of 4K tokens, using language as what MiniMax calls "the generalizable bridge" between context and output.
• H3-VAE is a rebuilt tokenizer that delivers a 4x gain in effective sequence length over the prior version, which is what makes native 2K generation computationally viable in the first place.
• H3-Omni Transformer training separates understanding and generation compute, lifting training throughput roughly 30% despite tripled sequence-length variance.
• In-Context Regeneration upscales the model's own low-resolution output instead of routing through a separate super-resolution model, which is how it preserves fine detail like on-screen text and branding.

(all four, source)

For a sense of scale beyond the 33B transformer itself: the full inference stack, including the Qwen3-VL conditioner, the transformer, and both VAEs, runs to roughly 69.2B parameters and about 134GB of on-disk weights for a single task partition (62.13 GiB for the text conditioner, 61.73 GiB for the transformer, 9.70 GiB for the visual VAE, and 0.56 GiB for the audio VAE) (source). Per Artificial Analysis, H3 currently ranks #1 in video editing, #2 in text-to-video, and #3 in image-to-video with audio among evaluated models as of early August 2026, a single third-party arena run rather than a MiniMax-published benchmark table (source).

What H3 Actually Ships Open vs What Stays API-Only (2K Upscale, Prompt IR)

Here's the part "open-weight" undersells: the checkpoint you can actually download is H3-Base only, released as FL2VA (text-to-audio-video plus first/last-frame control) and Ref2VA (reference-to-audio-video) variants, both natively generating at 768p (source). Both checkpoints are confirmed live on Hugging Face at MiniMaxAI/MiniMax-H3.

The Context Omni Representation stage and the Regenerate-2K upscale, the two pieces that produce the API's signature native 2K output, are not part of that open checkpoint. Kingy AI's hardware review put it plainly: "the official 2K workflow cannot run completely locally because Context-IR and Regenerate-2K are API-only at launch" (source). If your pipeline requires native 2K output today, that means the hosted API, not a local deployment.

That gap has a real cost comparison behind it. MiniMax prices the hosted API at roughly $0.13/second for 2K output (about $1.95 for a 15-second clip) and $0.08-0.09/second for 768p; the first 5 reference images per generation are free, then $0.04 each, and reference audio is free (source). At the 2K tier, that works out to roughly $7.80/minute, against approximately $20.16/minute for Kling 3.0 at 1080p and $22.45/minute for Dreamina Seedance 2.0 at 1080p (source). So even the part of H3 you can't self-host is priced well below the closed competition, which is worth knowing before you decide the open checkpoint's 768p ceiling is a dealbreaker.

GPU and VRAM Requirements for Self-Hosted Inference

Checkpoint download size scales with quantization:

(source: oflight.co.jp)

The pruned checkpoint gets its roughly 40% size reduction over standard INT8 by precomputing the AdaLN branch lookup tables ahead of time instead of calculating them at inference, since those ~13B AdaLN parameters don't need to be recomputed per request once cached (source).

Disk size tells you what to download; peak VRAM during an actual run is what determines whether a job finishes. Here's what's been measured directly, generating a 5-second, 50-step clip at 1344x768:

(all rows, source)

Treat the timing column as directional rather than a strict apples-to-apples comparison. Different configurations report pipeline latency, end-to-end time, or a denoise/decode split, and those aren't the same measurement. The VRAM column is the more actionable number for picking hardware.

For lower-VRAM local rigs: 24GB-class cards like the RTX 4090 reportedly work with the pruned INT8 + NVFP4 checkpoint plus layer-wise offloading, and the ComfyUI team has commented that 12GB "may work" given ample system RAM and fast storage, with no guarantee on speed or stability. Budget 64GB+ of system RAM for the pruned baseline, 128GB+ if you want headroom for the larger checkpoints (source).

Spheron's own catalog is datacenter-tier hardware, H100, H200, B200, B300, A100, and GH200, not RTX 4090 or RTX 5090. If your actual plan is to run the pruned checkpoint on a 24GB consumer card, that's a local workstation build rather than a cloud rental. On Spheron, the equivalent budget entry point is a single H100 PCIe (80GB), which swallows the ~42.5GB pruned checkpoint with plenty of headroom for activation memory. See the GPU and pricing section below for what that costs.

The License's Regional Deployment Restrictions and What They Mean for You

MiniMax H3's open weights ship under the MiniMax H3 Community License Agreement, effective August 2, 2026. Commercial use is permitted at no separate fee for organizations under $20 million in yearly revenue with attribution; above that threshold, you need separate written authorization from MiniMax before you can use the model commercially (source: license text).

The bigger catch is geographic. The license defines "Applicable Territory" as worldwide, excluding the European Union, the United Kingdom, the Republic of Korea, and the United States of America (source: license text). Using, modifying, distributing, or hosting the open weights, or their outputs, isn't licensed in those four regions without separately obtained authorization. MiniMax's hosted API remains available everywhere, including in the excluded regions; it's specifically local deployment of the open checkpoint that's gated.

The restriction traces back to an active lawsuit. Disney, Universal Pictures, and Warner Bros. Discovery sued MiniMax, alleging H3 was trained on their copyrighted material without authorization and can generate recognizable characters from franchises including Marvel and Star Wars. A US court rejected MiniMax's motion to dismiss the case in May 2026, letting the litigation proceed. Ryan Lee, MiniMax's Head of Developer Relations, confirmed the connection directly: "U.S. restrictions stem from a generative video copyright lawsuit involving major Hollywood studios" (source).

If you're a US-based developer who still wants to self-host, MiniMax says it will issue individual authorization to applicants who commit to a content compliance mechanism meeting US legal requirements, through its own licensing channel. That's a real path, but it's an application process, not a default right under the Community License.

Practically: if your organization is based and operating in the US, EU, UK, or South Korea, treat local H3 deployment as off the table until you've secured that individual authorization, and lean on the hosted API in the meantime. If you're outside those four regions, standard self-hosting applies, and this is one more reason to check where your GPU provider's compliance and data-residency options actually put your deployment; our guide to GPU cloud providers in India is a useful reference if you're evaluating capacity in a market this license doesn't touch. None of this is legal advice: read the full license text and confirm your situation with counsel before you deploy H3 commercially.

Deploy MiniMax H3 with ComfyUI on Spheron

ComfyUI added native MiniMax H3 support starting in version 0.30.0, with prebuilt Text-to-Video, Image-to-Video, and Reference-to-Video templates (source). That native support is why this is a comfortable ComfyUI deployment rather than a custom-node scramble; for a broader look at running ComfyUI on rented GPUs generally, see our ComfyUI on GPU cloud guide.

Step 1: Confirm Your License Status

Before provisioning anything, confirm your organization isn't operating in the US, EU, UK, or South Korea, or that you hold individual authorization from MiniMax. This is a five-minute check that saves you from a compliance problem after you've already paid for GPU time.

Step 2: Provision a GPU Node on Spheron

Pick your configuration from the VRAM tables above. A single H100 PCIe instance covers the pruned INT8 + NVFP4 checkpoint; a 4-GPU H100 or H200 configuration covers full BF16. SSH in using the SSH connection guide.

Step 3: Update ComfyUI and Download the Model Files

# Update ComfyUI to 0.30.0 or later for native H3 support
cd ComfyUI
git pull
pip install -r requirements.txt

The Comfy-Org/MiniMax-H3 repository hosts several precision tiers of the same diffusion model. To match the single H100 PCIe from Step 2, pull the pruned INT8 variant, not the full BF16 file, which alone runs ~66GB and won't leave headroom for activation memory on an 80GB card:

pip install huggingface_hub

# Diffusion model: pruned INT8 (~21GB) fits the single H100 PCIe from Step 2.
# Use fl2va for T2V/I2V, ref2va for R2V. Swap in minimax_h3_fl2va_bf16.safetensors
# (~66GB) only if you provisioned the multi-GPU BF16 configuration instead.
huggingface-cli download Comfy-Org/MiniMax-H3 \
  minimax_h3_fl2va_pruned_int8_convrot.safetensors \
  --local-dir ComfyUI/models/diffusion_models

# Text encoder (ships pre-quantized NVFP4/AWQ even in the native workflow)
huggingface-cli download Comfy-Org/MiniMax-H3 \
  qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors \
  --local-dir ComfyUI/models/text_encoders

# Video and audio VAEs
huggingface-cli download Comfy-Org/MiniMax-H3 \
  minimax_h3_video_vae_fp16.safetensors \
  --local-dir ComfyUI/models/vae

huggingface-cli download Comfy-Org/MiniMax-H3 \
  minimax_h3_audio_vae_fp32.safetensors \
  --local-dir ComfyUI/models/vae

Notice the text encoder ships pre-quantized to NVFP4/AWQ by default, not as an optional lightweight variant. That's consistent with the pruned-checkpoint strategy covered above: even the officially supported ComfyUI path leans on quantization to keep the footprint manageable.

Step 4: Load a Template and Generate

Launch ComfyUI, bound to localhost, and reach it through an SSH tunnel rather than exposing the port publicly:

ssh -L 8188:localhost:8188 user@your-server-ip

Open http://localhost:8188, go to the Template Library, select the Video section, and choose the T2V, I2V, or R2V MiniMax H3 template. Follow the pop-up to confirm the model paths, then queue a test generation before running anything at production volume.

Open-Source Video Generation Models in 2026: Where H3 Fits

Most flagship video models released in 2026 shipped API-only. Wan 2.5 (September 2025), Wan 2.6 (December 2025), and Wan 2.7 (April 2026) all followed that pattern from Alibaba's Tongyi Lab, with no public checkpoint on GitHub or Hugging Face for any of them. Wan 2.2, from July 2025, remains the last Wan release with publicly downloadable weights, and it carries no territorial restriction at all.

MiniMax H3 breaks the API-only trend in one sense, base weights are genuinely downloadable, but it introduces a different kind of gate: a license that's open in most of the world and closed by default in four major economies. That makes 2026's open-source video generation landscape split into three real categories rather than two: fully open with no restriction (Wan 2.2, HunyuanVideo, LTX-Video), open but region-gated (MiniMax H3), and closed API-only (Wan 2.5/2.6/2.7, Kling, Seedance). Which category matters to you depends entirely on where you and your deployment sit, not just on what the model card says.

For broader VRAM planning across this landscape, GPU cloud for video AI 2026 covers Wan 2.1 and HunyuanVideo sizing, and the image-to-video deployment guide covers LTX-Video and Wan 2.2 I2V specifically. If native synced audio is the feature that brought you to H3 in the first place, it's worth knowing that's not unique to video models: our guide to self-hosting open-source AI music generation covers YuE, ACE-Step, and Stable Audio Open for teams that need generated audio without the video attached.

MiniMax H3 vs Wan 2.2/2.5: Which Video Model to Self-Host Right Now

If your organization sits outside H3's four excluded regions, it's the more capable open checkpoint on paper: native audio in the same generation pass, longer clips, and a higher base resolution than Wan 2.2. It's also weeks old against Wan 2.2's year of production hardening and a wider base of tested ComfyUI workflows and community tooling. If you're inside the US, EU, UK, or South Korea, the license question settles it by default: H3 isn't a licensed local deployment option for you without individual authorization, which makes Wan 2.2 the practical open-source pick regardless of feature comparisons. See the Wan 2.5 deployment guide for the full Wan 2.2 ComfyUI and diffusers setup, and the Wan 2.7 guide for why that later release stayed API-only too.

Spheron GPU Configurations and Pricing for MiniMax H3

Pricing fetched from the Spheron API on 16 Aug 2026:

Pricing fluctuates based on GPU availability. The prices above are based on 16 Aug 2026 and may have changed. Check current GPU pricing → for live rates.

For the entry-level pruned checkpoint, H100 GPU rental on Spheron is the practical starting point: single-GPU, no tensor parallelism to configure, and enough headroom above the ~42.5GB checkpoint for activation memory during a 768p generation. For full BF16 quality, H200 GPU pricing on Spheron gives the better price-per-GB of the two datacenter options at spot rates, since 141GB per card absorbs the 94.3GB peak with more room than an H100's 80GB. Teams running FP8 at higher throughput should look at B300 availability on Spheron, currently spot-only in Spheron's catalog, where the 51.9GB peak per GPU leaves meaningful headroom even at 8-GPU scale.

MiniMax H3's open checkpoint gets you native-audio 768p video generation today, license permitting. Datacenter GPUs on Spheron give you the VRAM headroom to run it at full BF16 without babysitting offload configs.

Spheron H100 → | H200 for full BF16 → | View all GPU pricing →

Deploy MiniMax H3 on Spheron →