Aug 8, 2026

MiniMax H3 Open Weights: The 33B Video Model That Fits One RTX 5090 — and the License That Locks Out the US and EU

By RunAIHome Team · 12 min read

video-generationminimaxcomfyuigpuvramopen-weightshardware

TL;DR: MiniMax open-sourced H3 (Hailuo 3.0) on August 3, 2026 — a 33B omni-modal video model that generates 4–15 second clips with native stereo audio, and quantized builds really do run on one consumer GPU. The catch isn’t hardware: the Community License excludes the US, EU, UK, and South Korea from local deployment. For most of our readers, that’s the whole story.

	H3 local (where licensed)	H3 hosted API	Wan 2.x local (anywhere)
Best for	24–32GB GPUs outside the excluded territories	US/EU users who want H3 output legally	US/EU home labs that want local video today
Hardware / cost	RTX 5090 32GB (~175s per 10s clip, NVFP4)	$0 upfront, per-clip pricing	Runs from 8GB VRAM up; Apache 2.0
The catch	License excludes US, EU, UK, South Korea	Your prompts leave your machine	No native audio, 720p-class base output

Honest take: H3 is the most capable open-weight video model anyone has shipped, and the single-GPU quants are a genuine engineering feat. But if you’re in the US or EU — like roughly 85% of the people reading this — the license says the weights aren’t yours to run at home. Wan stays our local-video recommendation until that changes.

MiniMax released the H3 weights on Hugging Face on August 3, 2026, with ComfyUI support merged the same day. Within 24 hours the community had GGUF, INT4, and NVFP4 quants up, and a ComfyUI maintainer had the thing finishing renders on an RTX 3060. It’s the fastest open-weights-to-consumer-hardware pipeline we’ve tracked all year — faster than Kimi K3’s weights drop, and that one took days, not hours, to become runnable.

So this guide does two jobs: the hardware math for anyone who can legally run it, and the license fine print for everyone who can’t.

What H3 actually is

Early social posts called H3 everything from a hybrid SSM to an LLM. The official repo is clear: H3-Omni-Transformer is a 33B-parameter dense, single-stream transformer for video-plus-audio generation. About 13B of those parameters sit in AdaLN-related branches whose outputs can be precomputed and cached, so an inference-only deployment loads roughly 20B parameters of working weights. There’s no modality-specific attention or FFN structure — text, image, video, and audio all flow through the same stream, positioned by three-dimensional multimodal RoPE.

Output specs, from the model card:

Clip length: 4–15 seconds per generation, 24 FPS
Audio: native 32 kHz stereo, generated in sync — no separate TTS or foley pass
Resolution: 768p on the shorter side by default; the “2K” in the marketing comes from a separate H3-Regenerate-2K pass
Aspect ratios: 21:9, 16:9, 4:3, 1:1, 3:4, 9:16

The release ships as two task-specific checkpoints, both BF16 and CFG-distilled:

Checkpoint	What it does	Inputs
FL2VA	Text-to-video, image-to-video (first frame, last frame, or both)	0–2 images + prompt
Ref2VA	Reference-driven generation: identity, style, voice	Up to 9 images, 3 video clips, 3 audio clips (12 files max)

If you just want text-to-video or standard image-to-video, FL2VA is your download. Ref2VA is the one that clones a face and a voice from reference media, which is also — not coincidentally — the checkpoint that makes regulators nervous.

The license: read this before you download 40GB

H3 ships under the MiniMax H3 Community License Agreement, not Apache 2.0 or MIT. Three clauses matter:

Territory exclusion. The license’s “Applicable Territory” excludes the United States, European Union, United Kingdom, and South Korea. Local deployment in those regions requires separately obtained authorization from MiniMax (there’s an application form in the repo). Coverage of the exclusion quotes MiniMax attributing it to the regulatory climate around generative video — the EU AI Act, evolving UK and Korean rules, and active US copyright litigation.
Revenue threshold. Companies over $20 million in annual revenue need separate commercial authorization regardless of territory.
Attribution. Commercial products built on H3 must display “MiniMax H3” in the UI.

To be precise about what this is: a license restriction, not a technical block. The weights download the same everywhere, and nobody’s GPU checks a passport. It’s a civil-contract line, and how it would ever be enforced against a hobbyist is anyone’s guess. But we’re not going to tell you to ignore a license — we made the same call on Kimi K3’s fine print, and video models carry more legal heat than text models right now, a theme we covered in the right-to-compute legislation piece. The hosted Hailuo API is a different legal relationship and remains available globally; that’s the compliant route for excluded territories.

The rest of this guide is the hardware reality for readers in the ~190 countries where local deployment is licensed — and for everyone else’s reference the day the territory list changes.

The download, by size

Full-precision deployment of a checkpoint plus its text encoder and VAEs comes to roughly 115GB — that’s multi-GPU or server territory. Nobody at home runs that. What you actually download, from smallest to largest:

Build	Disk size	Source
GGUF Q2 UNet (smallest community quant)	8.49GB (+ text encoder + VAE)	community repos
GGUF Q3_K_S/M UNet	~15.6GB (+ encoder + VAE)	community repos
GGUF Q5_K_M UNet	~23.9GB (+ encoder + VAE)	community repos
NVFP4 pruned, complete 4-file set	31.7GB total	community (RTX 50-series only)
Comfy-Org optimized workflow set (pruned INT8)	~42.5GB working set	Comfy-Org/MiniMax-H3
BF16 single checkpoint, smallest form	~21GB per checkpoint (model only)	MiniMaxAI/MiniMax-H3

Two things the table hides. First, the text encoder is enormous — community GGUF builds pair the UNet with a text encoder running 14.6–27.1GB depending on quant, plus about 5.8GB of visual and audio VAEs. The UNet number alone undersells the footprint; budget for the full stack. Second, the Comfy-Org repackage hosts bf16, INT8, and pruned-INT8 diffusion variants plus bf16/INT8/NVFP4-AWQ text encoders — the whole repo is around 343GB, so download the specific files your workflow needs, not the repo.

Real numbers on real GPUs

These are the verified data points, not extrapolations:

RTX 5090 (32GB) — the sweet spot. An independent benchmark ran the pruned NVFP4 build (four files, 31.7GB on disk) on a single RTX 5090: a 243-frame, 10.1-second, 864×480 clip with audio at 10 steps took 175 seconds, peaking at 26,914 MiB VRAM. Sub-3-minutes per clip is workable for iteration, and 26.9GB peak explains why this config exists only for 32GB cards. NVFP4 needs Blackwell — if you’re wondering why that format keeps showing up on RTX 50-series, we broke it down in the ComfyUI NVFP4 guide.

RTX 4090 (24GB) — possible, with system RAM doing heavy lifting. Community tests with FP8 diffusion plus a NVFP4/AWQ text encoder saw total memory peak around 46GB, which means substantial spillover into system RAM even on a 24GB card. On a 128GB-RAM rig with SageAttention enabled, 12-second HD clips averaged 10–13 minutes. That’s render-farm patience, not iteration speed. The fix that got it stable: pruned INT8 or FP8 weights, SageAttention, and ComfyUI’s dynamic offloading — trying to load the raw BF16 checkpoint on a 24GB card just OOMs.

RTX 3060 (12GB) — proof of concept. A ComfyUI maintainer completed the 0.4-megapixel, 124-frame, 20-step template workflow on an RTX 3060 12GB with 32GB of system RAM and fast NVMe, leaning entirely on dynamic VRAM offloading. It finishes. Nobody reported it being pleasant. If your card is in the 8–16GB class, the Q2–Q3 GGUFs (8.49–15.6GB UNet) are the realistic entry point, and you should expect minutes-per-second-of-video render times.

No verified Mac numbers yet — the ComfyUI path is CUDA-first at launch, and we’ll update when credible Apple Silicon runs surface.

GPU tier summary

GPU class	Build to use	Expect
RTX 5090 32GB	NVFP4 pruned (31.7GB)	~175s per 10s 864×480 clip, 26.9GB peak VRAM
RTX 4090 / 3090 24GB	FP8 / pruned INT8 + offload	10–13 min per 12s HD clip; wants 64–128GB RAM
16GB (5080, 5060 Ti 16GB)	GGUF Q3–Q4 + offload	Untested-but-plausible; heavy offload, long renders
12GB (3060, 4070)	GGUF Q2–Q3, 0.4MP workflow	Confirmed to complete; 32GB+ RAM and NVMe required

System RAM is the silent requirement in every row below the first. With a 42.5GB working set against a 24GB card, ComfyUI is streaming weights constantly — 32GB RAM is the floor, 64GB+ is comfortable, and slow storage will hurt more than a slow GPU. Our system RAM guide logic applies double here.

Running it: ComfyUI setup

Native support landed in ComfyUI core the day of release (PR #15224), so there’s no custom node hunt for the official workflows. You need ComfyUI 0.30.0 or later:

cd ComfyUI && git pull
python main.py
# startup banner should report:
# ComfyUI version: 0.30.0 (or later)

From there, the H3 template ships in the built-in Template Library (Workflow → Browse Templates → Video), which pulls the Comfy-Org file set for you. GGUF builds instead load through the usual ComfyUI-GGUF custom node with the UNet, text encoder, and VAE files placed in their standard folders.

The most common failure we’ve seen reported is the classic one: “torch.OutOfMemoryError: Allocation on device” partway through sampling on 24GB cards running FP8 with a too-ambitious resolution. The working fix, per community testing: drop to the pruned INT8/NVFP4 diffusion model, enable SageAttention, and let ComfyUI’s offloading do its job rather than forcing everything resident — the 46GB peak has to live somewhere, and on a 24GB card that somewhere is system RAM. Same failure mode we cataloged in the CUDA OOM fix guide.

H3 vs Wan: the question that actually matters at home

For US and EU readers, the practical comparison isn’t H3-local vs H3-API — it’s H3 vs Wan, the Apache 2.0 video family that has carried local video generation all year.

H3 wins on capability, and it isn’t close: native synchronized stereo audio, omni-reference identity and voice control, 15-second clips, and a 2K regenerate pass. Wan generates silent video and tops out shorter. But Wan runs on an 8GB card without a lawyer, its license has no territory list, and its ecosystem (LoRAs, ControlNets, speedup tricks) is a year deep. If your use case is B-roll, product shots, or animation experiments in an excluded territory, Wan remains the answer. If you’re in a licensed territory with 24GB+, H3 is now the most capable thing you can run in your own case.

And if you need H3-class output from the US or EU today, the compliant options are the hosted Hailuo API, or renting compute and using the API-side products — a RunPod instance is also the cheap way to benchmark the open builds’ output quality before deciding whether any of this matters for your workflow.

Honest take

The engineering story is real: 115GB of full-precision model compressed to a 31.7GB set that renders a 10-second clip with sound in under three minutes on one consumer card, with day-0 support in the tool everyone already uses. That’s the open-weight video future arriving on schedule.

The license story undercuts it. A model whose weights 85% of our audience cannot legally deploy at home is, for that audience, a hosted product with extra steps. MiniMax was at least explicit about why — video generation sits in regulatory crosshairs that text models have so far dodged — and the exclusion list reads like a map of pending AI legislation. Watch for one of two things: MiniMax granting territory authorizations (the form exists), or a genuinely open competitor matching the native-audio trick. Either one changes our recommendation overnight. Until then: Wan at home, H3 in the cloud, and keep your 24GB card — it stays useful either way.

FAQ

Is MiniMax H3 an LLM? No. Despite the “H3” name colliding with SSM-hybrid language models like MiniMax’s own M-series text models (see our MiniMax M3 guide), H3 is a 33B dense transformer for video-and-audio generation — it’s the open-weights release of Hailuo 3.0.

Can I legally run H3 at home in the US or EU? The Community License excludes the US, EU, UK, and South Korea from local deployment without separate authorization from MiniMax. It’s a license term, not a technical block, and there’s an authorization application in the repo. The hosted API remains available in all regions.

What’s the minimum GPU that actually works? The confirmed floor is an RTX 3060 12GB with 32GB system RAM and NVMe storage, running the 0.4MP template workflow with dynamic offloading. Comfortable starts at 24GB, and the single-card sweet spot is the RTX 5090 with the 31.7GB NVFP4 build.

Does the 33B model fit in 24GB of VRAM? Not resident — community tests peaked around 46GB total with FP8 weights, so a 24GB card depends on ComfyUI streaming weights from system RAM. It works, at 10–13 minutes per 12-second HD clip.

Is the audio really generated with the video? Yes — 32 kHz stereo, generated in the same pass, synchronized to the visuals. No post-hoc TTS step. It’s the first open-weight model to ship this.

Recommended Gear

RTX 5090 — the only consumer card that runs H3’s NVFP4 build fully resident (26.9GB peak)
RTX 4090 — 24GB + big system RAM gets you there with offloading, slowly
RTX 3060 12GB — the confirmed budget floor for the low-res workflow

Sources

Last updated August 8, 2026. Prices, specs, and license terms change; verify current terms before downloading or purchasing.

Was this article helpful?