MiniMax H3 Is Here, Uncensored and Wild: The Open-Source Video Model That Generates Audio Too

The open-source video generation landscape just got a caffeine enema. MiniMax dropped the weights for H3, their 33-billion-parameter omni-modal monster, and it doesn’t just generate video, it generates synchronized stereo audio in the same pass. The early results from the community are nothing short of absurd. We are talking 2K resolution, up to 15 seconds of content, and according to early testers on RTX 5090s, it’s “fully uncensored” and follows prompts better than anything we’ve seen before.

This isn’t just another incremental model release. This is a paradigm shift in how we think about generating content. Forget stitching together a text-to-video model, an image-to-video model, and a text-to-speech model. H3 eats all your inputs, text, images, video, and audio, as a single context and vomits out a finished product. The only catch? The license is a doozy.

The One Model to Rule Them All

Previous video generation stacks were a mess of specialized experts. You had a model for text-to-video, another for image-to-video, another for first-and-last-frame interpolation, and a whole separate ecosystem for audio generation and editing. MiniMax looked at this Frankenstein architecture and decided to burn it down.

H3 is built on a simple, radical philosophy: unify everything. The core architecture is the H3-Omni-Transformer, a 33B-parameter dense single-stream Transformer. Critically, it uses a single set of weights for understanding and generation. There are no gated experts, no modality-specific attention layers hidden in the main trunk. The model uses Qwen3-VL-32B as a text encoder, feeding hidden states into the main transformer. This design choice is a deliberate departure from the Hailuo-02 architecture, which MiniMax deemed “unnecessary complexity” for a model built around task generalization.

The result is a system that can take a prompt like “Reference the Hitchcock camera movement from Video 1, have the character in Image 2 sing, with the vocals matching Audio 3” and just… do it. The H3-Context-IR module (which remains a closed API) acts as the brain, distilling roughly 100K tokens of raw input into an average of 4K tokens of structured instruction. This is where the magic of complex multimodal understanding happens.

Technical Deep Dive: The VAE and the Transformer

You can’t just throw 100,000 tokens of video and audio at a Transformer and expect it to work. The secret sauce is the H3-VAE, a completely rebuilt tokenizer that delivers a claimed 4× gain in effective sequence length. This high compression ratio is the lynchpin that makes native 2K resolution economically viable.

The H3-Omni-Transformer itself is a 33B-parameter model, but about 13B of those are in AdaLN (Adaptive Layer Normalization) branches. Because these can be precomputed and cached, the effective inference load is closer to 20B parameters. The model uses 3D Multimodal Rotary Position Embeddings (MM-RoPE) to represent temporal and spatial relationships (t, h, w), allowing it to understand how a character in frame 1 moves to frame 40.

For the 2K resolution, MiniMax doesn’t use a traditional super-resolution module. Instead, it uses a technique called H3-Regenerate-2K. The model feeds its own low-resolution output back into itself along with the original context to regenerate the video at a higher resolution. This in-context approach recovers fine details that standard upscalers can only guess at.

The “Uncensored” Reality

The Reddit community response, particularly from user FinBenton who tested it on an RTX 5090, is the most telling part of this story. The claims read like a wishlist for local AI video generation:

• Fully uncensored: Straight out of the box, no jailbreaks required. The model will generate explicit content, violence, and potentially sensitive political themes without the moralizing safety filters that plague closed-source models like Sora or Veo.
• Better prompt following: Complex, multi-modal prompts are followed to the letter. The model understands the relationships between inputs, not just the inputs themselves.
• Wild audio generation: It doesn’t just do dialogue. It does ambient sounds, Foley effects, and bizarre combinations “to the T.” Any sound, any action.

A sentiment echoed on developer forums is that “This will be the new wan2.2 for a very long time.” The community is already preparing. ComfyUI’s repackaged model files are live, and workflows for text-to-video (T2V) and reference-to-video (R2V) are ready to go.

The License Problem That Can’t Be Ignored

Here is where the party gets a bucket of cold water. The MiniMax H3 Community License Agreement is not your typical permissive open-source license. It is geographically restricted, explicitly excluding the European Union, the United Kingdom, the Republic of Korea, and the United States of America.

“Excluded Territories” means the European Union, the United Kingdom, the Republic of Korea and the United States of America.

This means if you are a developer in New York, London, or Berlin, you are legally prohibited from using the model without obtaining “written permission” from MiniMax. MiniMax has acknowledged this, citing “Hollywood lawyers” and regulatory concerns in their Q&A. They do offer exemptions, but the friction is significant.

This creates a bizarre dynamic. The model that is creating the most buzz is essentially legal in China and a handful of other territories, but illegal in the West. For those in the excluded territories, the API remains the only path. The pay-as-you-go pricing for the API is $0.13 per second for 2K video, with the first 5 reference images free.

This restriction puts H3 in a category similar to other open-weight AI models with restricted multimodal capabilities like Mistral’s Voxtral, which offers impressive performance but keeps its best features behind an API.

Hardware and Deployment: The 5090 Factor

You aren’t running this on a college laptop. The full 33B model in BF16 requires substantial VRAM. Early community reports, however, are promising for those with high-end hardware.

Deployment with SGLang is straightforward:

# For First/Last-Frame mode
sglang serve \
  --model-path MiniMaxAI/MiniMax-H3 \
  --num-gpus 4 \
  --ulysses-degree 4 \
  --performance-mode speed \
  --host 0.0.0.0 \
  --port 30010 \
  --model-variant fl2va

The fact that ComfyUI, the standard-bearer for visual AI workflows, had native support within hours of the weight release is a testament to the model’s impact. You can now build complex workflows that ingest a brand image, a motion reference video, and an audio clip to produce a finished ad spot in a single node graph.

This shift toward running complex models locally is part of a larger trend. As noted in our coverage of running multimodal models locally on accessible hardware, the era of mandatory cloud inference is ending.

What This Means for the Industry

The implications of H3 are staggering for content creation, but they are also deeply unsettling. The democratization of high-quality, uncensored video generation is a double-edged sword.

H3 is going to be a bonanza for budget-constrained studios and individual creators. But it is also going to be a nightmare for anyone trying to enforce consent, copyright, or political stability. The “uncensored” nature of the model is a feature for creativity and a bug for societal safety.

We are seeing a similar dynamic play out in the world of open-source multimodal world models with video-like generation, where the ability to simulate environments is outpacing our ability to govern their use.

MiniMax H3 is a technical masterpiece that arrived with a legal anchor wrapped around its leg. The model itself is clearly a leap forward, perhaps the most significant open release in the video generation space since Stable Diffusion disrupted image generation. The weight release (for those who can legally access it) signals a move towards a more open ecosystem, which is good for innovation.

But for the vast majority of the Western developer community, this is a tease. You can look, but you can’t touch without getting your legal team involved. For now, the API is the only ethical (and legal) path forward for many of us.

If you are in a permitted territory, or if you want to take the risk and run it anyway, you are about to witness what the next generation of AI video looks like. For everyone else, download the weights from a mirror, run it locally, and wait for the legal landscape to catch up. The technology is ready, the law is not.