MiniMax H3 Open Weights Are Out — And There Are Three Catches

The wait is over. MiniMaxAI/MiniMax-H3 went live on Hugging Face on August 3, 2026, roughly a day after the countdown expired without a release. It picked up over 1,400 likes within hours.

We checked the repository directly rather than working from announcement threads, and three things in it differ sharply from how the launch was marketed. If you are planning to run H3 locally, read the license section first — for a lot of readers it is decisive.

Status as of August 4, 2026

Thing	Status
H3 via hosted API (fal, ComfyUI, Runway, Leonardo, Topview)	✅ Live since July 31, globally available
Official open weights on Hugging Face	✅ Published August 3
Usable under the license in the US / UK / EU / South Korea	❌ Excluded — see below
Base model output resolution	⚠️ 768p; 2K needs a second pass

Catch 1: The License Excludes the US, UK, EU and South Korea

This is the most important and least reported detail. H3 ships under the MiniMax H3 Community License Agreement, not Apache or MIT. The grant is limited to an "Applicable Territory", defined as:

"Applicable Territory" means worldwide, excluding the Excluded Territories. "Excluded Territories" means the European Union, the United Kingdom, the Republic of Korea and the United States of America.

So if you are in the US, UK, EU or South Korea, the community licence does not grant you the right to use, reproduce, distribute or modify the weights. MiniMax invites people in those regions to contact them for a separate licence "granted based on robust controls and guardrails".

MiniMax published a docs/QA-about-License.md explaining the reasoning, and to their credit it is direct about it:

Regions such as the EU, UK, South Korea, and the US are currently developing or enforcing AI-related regulations that may have specific implications for generative video models, especially around areas such as likeness generation, copyright, content safety, and responsible deployment.

They cite the EU AI Act's enforcement, regulatory uncertainty in the UK and South Korea, and — notably — that MiniMax is "involved in ongoing copyright-related legal proceedings specifically concerning generative video AI" in the US. Their stated position is that the limitation means "not yet, not not ever."

Important distinction: this restricts the weights, not the hosted API. The API remains globally available, because MiniMax controls the serving infrastructure and can enforce safeguards there. If you are in an excluded territory, fal, Runway, Leonardo, ComfyUI's partner nodes and Hailuo itself are all still open to you. You just cannot legally run the downloaded model yourself.

Catch 2: "Native 2K" Is a Two-Stage Process

The launch messaging emphasised native 2K. The repository is more precise. From the model card:

Output resolution: Supports various resolution dimensions. The shorter side is set to 768 pixels by default. 2K generation can be achieved with H3-Regenerate-2K.

H3 is actually a three-module system:

• H3-Context-IR — takes your messy multimodal input and refines it into a "Context Intermediate Representation" the generator consumes.
• H3-Base — generates audio and video at 768p.
• H3-Regenerate-2K — feeds the 768p result plus the original context back through H3 to regenerate at 2K.

So 2K is a regeneration pass, not a native output mode of the base model. That is not a bad design — regenerating with the original context available produces better detail than upscaling — but it means your local pipeline is two generation passes, not one, if you want 2K.

The rest of the confirmed spec: 4–15 second outputs, 24 FPS, 32 kHz stereo audio, aspect ratios from 21:9 through 9:16, and stable dialogue support in 11 languages (Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, Spanish).

Catch 3: It Is Very Large

Before release, <a href="https://x.com/cheatyyyy/status/2083863807211049110" rel="nofollow" target="_blank">cheatyyyy</a> read the open PRs adding H3 support to ComfyUI and diffusers, and reported "33B for the main DiT and a pruned 20b variant", ending with "idk if this is going to run on a 3090 bros".

The 33B figure checks out — the transformer folder is 66.3GB, which is 33B parameters at bf16. The pruned 20B variant is not in the released repository. It may still arrive; it did not ship on day one.

The repository totals roughly 499GB, because it contains two complete bundles (FL2VA and Ref2VA) alongside the shared components. A single usable pipeline is approximately:

Component	Size (bf16)
Text encoder (Qwen3-VL-32B)	66.7GB
Transformer (DiT, 33B)	66.3GB
Video VAE	10.4GB
Audio VAE	0.6GB
Total	~144GB

So the answer to "will it run on a 3090" is no, not at bf16, and not close.

Update: community quantisation closed that gap within days. The DeepBeepMeep/MiniMax-H3 repo now ships pruned ~20B int8 builds at around 21GB alongside GGUF text encoders, bringing a minimum working set down to roughly 35GB. We tested it end to end on a 12GB RTX 3060 — it runs, at 25 minutes per 5-second clip, with VRAM peaking at just 4.7GB, because the real bottleneck turns out to be system RAM rather than the GPU. Full walkthrough and the six errors you will hit along the way: running MiniMax H3 locally on an RTX 3060.

The Two Variants

The release ships two base models with different input modes:

• H3-Base-FL2VA (first-and-last-frame): zero images gives text-to-video, one image gives first- or last-frame conditioning, two images generates the transition between them.
• H3-Base-Ref2VA (omni-reference): up to 9 images, 3 video clips and 3 audio clips, with a hard cap of 12 files total. Each video or audio clip must be 2–15 seconds, and total duration cannot exceed 15 seconds. Audio cannot be the sole input — it must accompany an image or video.

Those reference ceilings match what fal published at launch, and they are the reason multi-character consistency worked so well for creators in week one.

The Part That Matters Most for Prompting

Buried in the model card is a line worth taking seriously:

H3-Context-IR is critical to the quality of the final output, so we strongly recommend incorporating it into your generation pipeline or following the "Prompting Guidance" to build your own context-processing system.

In other words, MiniMax is saying that a large share of output quality comes from how your input is preprocessed into structured context — before the generator ever runs. They ship two prompt-writing guides in the repo (docs/VIDEO_PROMPT_WRITING_GUIDE_base_en.md and ..._ref_en.md) for exactly this.

This validates what creators converged on in the first week without knowing the architecture: the prompts that worked were the ones that pre-structured the request — explicit camera sequencing, a named closing frame, labelled blocks for style and lighting. We broke those patterns down, including a continuous-shot template that spread between creators, in our MiniMax H3 launch prompts writeup. The baseline structure for writing to this model family is in the Hailuo H3 prompt guide.

If you are still using the hosted API — which, given the licence, many readers will be — <a href="https://x.com/sidharth_raja/status/2084063524452753424" rel="nofollow" target="_blank">sidharth_raja</a> noted a real price gap between providers: fal at $0.26/sec versus OpenRouter at $0.13/sec. On a 15-second clip that is $3.90 against $1.95.

Keep Track of What Creators Are Running

We collect AI video prompts from X continuously and sort them by model, so H3 and Hailuo examples accumulate as people publish them — browse them on trending prompts.

If you want to go the other direction and recover the prompt behind a clip you have seen, that is what our video to prompt tool does: paste a link or upload the video and it returns a structured prompt covering scene, motion and camera work.