2K, stereo sound, and use cases from ads to e-commerce. All official samples are below.

Official samples fall into two groups: ones showing raw capability (2K quality, native stereo) and ones showing use cases (intros, product pages, posters, ads). Headphones are recommended for the stereo clips — the model generates sound and picture together, so there's no post-production audio work.

2K costs less than half of what rivals charge for 1080p.

The official rate card is straightforward: $0.13/second for 2K, $0.08/second for 768p. The release post says "2K pricing is under a third of mainstream models," and a third-party pricing comparison lays that out concretely:

2K resolution at less than half the price of a rival's 1080p. The low cost and the versatility come from the same design decision.

It used to be one model per task. H3 tore that wall down.

Let's start with the old world. Image generation was split by task: text-to-image, editing, subject reference, style reference — each had its own specialist model. In audio, voice, sound effects, and music were three separate research tracks. Video was the most fragmented: text-to-video, image-to-video, first/last frame, subject reference, motion reference, editing — all separate. Users had to shuffle assets between tools, and models only ever saw their own narrow slice of data. Task boundaries capped generalization.

H3's first principle: tear that wall down. From the first day of pre-training, text-to-image, text-to-video, and text-to-audio train together. Video comes with sound natively; voice, effects, and music aren't siloed. "Reference" and "edit" are no longer fixed task lists but relationships described in natural language. That capability trains on real, natural data entirely, and it scales with data volume. Language becomes the bridge for task generalization.

Four techniques, all in service of task generalization.

1. Describing assets down to the "relationship" level

Generative models learn from "asset + description" pairs. H3 takes description writing down to the relationship level: not just describing the target video, but also the relationships between reference assets, the relationship between references and the target, and the audio-visual correspondence across multiple shots. MiniMax calls this Contextual Omni Representation. For this, they built a dedicated understanding pipeline: each training asset averages around 100k tokens of inference (roughly the length of a short novel), distilled into a precise description of about 4k tokens. H3's ability to follow instructions — "you say it, it gets it" — is rooted here.

2. H3-VAE: A complete rewrite of the video compressor

The new compressor reduces spatial dimensions by 16x and temporal by 4x. After patching into the model, the effective spatial compression is 32x. This yields 4x the effective sequence length, dramatically cutting training and inference costs. That's what makes 2K affordable. On the audio side, the same compressor processes each stereo channel separately before merging, so stereo is inherent, at a 32kHz sample rate.

3. Generation core: 33B parameters, one network, modalities stay together

The generation core is a 33B-parameter dense, single-stream Transformer. Attention and feed-forward layers don't distinguish between modalities. Modal differences live only in the input/output layers and the AdaLN branches — side pathways that send a "what type of content is being generated right now" control signal to each layer. Roughly 13B of the 33B parameters are in the AdaLN branches, and these can be pre-computed and cached, so they don't need to be loaded during inference:

The model card also reveals two engineering choices. First, the text and vision encoders directly use the full pre-trained weights of Alibaba's open-source Qwen3-VL-32B, feeding the hidden states from its 50th layer into the generation core — domestic open-source models using each other as components. Second, the architectural advantages accumulated in Hailuo 02 were entirely abandoned. That architecture was "redundant complexity for a model designed around task generalization." The thinking: architectures yield to model definitions. On the training side, "understanding" and "generation" compute loads are scheduled separately, boosting end-to-end training throughput by nearly 30%. Sparse attention for long sequences is natively trained, but the initial release only includes dense attention inference. The sparse implementation will be released separately later.

Official H3-Base architecture: text goes through the H3-Encoder (Qwen3-VL-32B), images go through both the encoder and a vision VAE, audio goes through an audio VAE. Everything is packed into a single sequence for the Omni Transformer, then decoded separately into video and stereo audio. Source: HuggingFace model card

4. 2K is re-generated with the original assets in hand

The traditional approach trains a dedicated super-resolution module to upscale 768p footage. Blurry small text and fine textures are mostly guessed. H3's 2K path is different: the base model carries the original reference assets and re-generates its own 768p draft from scratch. Because the original context is still available, small text and details can be genuinely recovered. Again, task generalization in action — upscaling is just another form of "generation."

Only the middle of the three-stage pipeline is open. Pre-processing and 2K require the official API.

The full H3 system is a three-stage pipeline, and only the middle stage has its weights released:

Stage by stage. The front end, Context-IR, handles your jumbled input — a few images, clips, a sentence — and turns it into structured descriptions for the generation model. It's "crucial to final output quality" (the model card's words), but it's not open-sourced — it relies on multi-stage workflows and multiple hosted models. Only the API is provided, along with two prompt-writing guides for the community to build this layer themselves. The middle, H3-Base, is the genuinely open part: two sets of weights. FL2VA handles text-to-video and first/last frame. Ref2VA handles full-modality reference. Both are CFG-distilled, BF16, outputting 768p. The back end, Regenerate-2K, "is complex, will release when ready" — API-only for now.

Both API ends have clear pricing. Pre-processing (Context-IR) is billed per token: $0.90 per million input tokens, $3.60 per million output. The 768p-to-2K regeneration is billed per second of output: $0.05/second, plus you pay the input fee for your original assets again. Running 768p locally costs nothing. For official-grade 2K, expect roughly $3 extra per minute.

The deployment ecosystem was complete at launch: SGLang, vLLM, diffusers, and ComfyUI all supported on day one. The official example command uses 4 GPUs; exact VRAM requirements aren't specified. Native ComfyUI support went live the same day as the weights, organizing H3's capabilities into five task types: text-to-video, image-to-video, first/last-frame control, reference-to-video, and in-place editing (modifying existing shots directly, avoiding full regeneration). Workflow templates and tutorials are ready to download.

Official system overview: the division of labor between Context-IR, H3-Base, and Regenerate-2K. Source: HuggingFace model card

Available worldwide, except the EU, UK, South Korea, and the US.

Downloadable weights don't mean usable everywhere — and the direction is the opposite of what most people assume. H3 uses MiniMax's own community license (MiniMax H3 Community License). The territory clause reads: "worldwide, excluding the following regions." Those are the EU, the UK, South Korea, and the US. Mainland China is in the licensed territory. Download and use. For the four excluded regions, you need to apply; MiniMax reviews the deployment scenario and compliance before granting a separate license.

The logic behind the exclusion list: where regulation is tightening, weights aren't provided first. The EU's AI Act is already in enforcement. The UK and South Korea have similar uncertainty. The US has an extra layer — the license FAQ explicitly points to a lawsuit MiniMax is involved in:

"In the US, AI regulation is still changing rapidly, and MiniMax itself is party to copyright-related litigation specifically targeting generative video AI."

MiniMax H3 License FAQ (translated by xiaohu.ai)

The FAQ ends with a promise: the current scope reflects the regulatory reality today. "It's 'not yet,' not 'never.'" Future changes will be announced explicitly, not silently updated.

No technical report yet. The claims are on the books, pending review.

Three things are in progress: a full technical report (the page just says "soon"); the next H-series model integrating MiniMax's own M-series language model capabilities; and continued scaling of model size and visual fidelity. The vendor also listed its own weaknesses at the end of the release post: multimodal understanding still has significant room for improvement, and visual details in some scenarios still need work.

Until the technical report and third-party evaluations are out, phrases like "production-ready quality" and "excellent instruction following," as well as the unnamed "mainstream model" in the pricing comparison, are all the vendor's own claims. The sample clips are above. You can check them yourself.

Output specs and pricing are fully transparent.

Half the reason 2K is affordable is the compressor. H3 rewrote the component that packs video into something the model can process, cutting spatial dimensions to 1/32. The model has far less to compute. The other half is a different approach to upscaling.

The full H3 system is a three-stage pipeline. This release only includes weights for the middle stage.

Running 768p locally costs zero API. For MiniMax-grade 2K: pre-processing is billed per token, regeneration is $0.05/second, so a minute of footage costs about $3 extra. The deployment ecosystem was ready at launch: ComfyUI, vLLM, SGLang, and diffusers all supported on day one. The example command uses 4 GPUs.

Downloadable weights don't mean usable everywhere — and the direction is the opposite of what most people assume. H3 uses a custom community license. The territory clause reads: "worldwide, excluding the following regions."

From the license FAQ: where regulation is tightening, weights aren't provided first. The EU's AI Act is in enforcement. The UK and South Korea have similar uncertainty. The US has an extra layer — the FAQ points to a copyright lawsuit MiniMax is involved in, specifically targeting generative video AI. The wording: "It's 'not yet,' not 'never.'"