--- source: 'https://howaiworks.ai/blog/minimax-h3-open-weights-video-model' section: blog title: 'MiniMax H3: 2K Video With Native Audio, Weights Shipped' description: >- MiniMax announced H3 on July 31, 2026: one model taking text, images, video and audio, out to 2K video with stereo sound. The weights shipped August 3. date: '2026-08-02' lastUpdated: '2026-09-06' author: HowAIWorks Team tags: - MiniMax - H3 - Video Generation - Multimodal AI - Open Weights - AI Video - Chinese AI - Text-to-Video - AI News newsSource: MiniMax breakingNews: false readingTime: 5 minutes isReviewed: false --- # MiniMax H3: 2K Video With Native Audio, Weights Shipped > MiniMax announced H3 on July 31, 2026: one model taking text, images, video and audio, out to 2K video with stereo sound. The weights shipped August 3. ## Introduction > **Update — September 6, 2026.** The weights shipped. `MiniMaxAI/MiniMax-H3` went live on Hugging Face on **August 3, 2026** — 33 billion parameters, tagged image-text-to-video, two base checkpoints (FL2VA and Ref2VA), and roughly 5 million downloads over the past month. Two qualifiers change the picture this article drew. First, `H3-Context-IR` and `H3-Regenerate-2K` were **not** released, so a local install tops out at a 768-pixel short edge — the 2K in the headline stays behind MiniMax's API. Second, the **MiniMax H3 Community License** defines "Excluded Territories" as the European Union, the United Kingdom, the Republic of Korea and the United States, and separately requires written authorization from MiniMax for anyone whose products generate more than $20 million a year. For a large share of this site's readers, "open weights" here does not mean "you may run it." > > The rate card exists now too. `MiniMax-H3` bills per second of output: **$0.13/s at 2K** ($7.80 a minute) and $0.08/s at 768P. A second variant, `MiniMax-H3-Max`, runs $0.08/s at 768P and $0.05/s at 480P, text-to-video and image-to-video only. (fal markets an "H3 Max" that it says its own research team post-trained; MiniMax states no provenance for its H3-Max rows, so treat the two as related but not confirmed identical.) For scale, Artificial Analysis's cost comparison — its own derivation, not vendor rate cards — puts [Kling 3.0](https://howaiworks.ai/models/kling) near $20.16 a minute at 1080p and Dreamina Seedance 2.0 near $22.45. The maintained version of all of this lives on our [MiniMax H3 (Hailuo 3.0)](https://howaiworks.ai/models/hailuo) model page. Everything below is the August 2 reporting, left as written. MiniMax announced **H3** on July 31, 2026: a single multimodal generation model that takes text, images, video and audio as input and returns video with native stereo sound, up to 15 seconds at 2K resolution. The Shanghai company — which listed in Hong Kong in January 2026 — says it will publish the model weights "in the coming days," which would make H3 the first flagship-tier open-weight video model. That last part is why the release matters beyond MiniMax's own catalog. Open weights have been the norm for language models and the exception for video, where the leading systems from ByteDance, Kuaishou, OpenAI and Google are all closed. As of August 2, 2026, no H3 checkpoint had appeared on MiniMax's Hugging Face or GitHub organizations, so the claim remains a promise rather than a download. ## What H3 Generates The pitch is a single prompt that mixes modalities. MiniMax's own example asks the model to reference a camera movement from one video, have a character from an image sing, and match the vocals to a supplied audio track — one instruction, three reference types, one output. Reporting on the launch puts the ceiling at nine images, three video clips and three audio tracks in a single generation. The output is 2K video with jointly generated stereo audio — voice, sound effects and music, not a soundtrack bolted on afterwards. MiniMax highlights instruction following, accurate text and brand rendering (the failure mode that makes most video models useless for advertising), and video-to-video motion transfer. The showcased use cases are commercial: film titles, product pages, animated posters, e-commerce ads. ## One Model Instead of a Task Menu The architectural argument is the more interesting half of the announcement. MiniMax's earlier [Hailuo](https://howaiworks.ai/models/hailuo) generations were a menu of specialists — text-to-image, image editing, subject reference, style reference, text-to-video, image-to-video, separate audio models for speech, effects and music, each with its own interface and its own constraints. H3 collapses those into one pretraining objective. Reference and editing relationships are expressed in natural language rather than selected from a fixed task list, which is what lets the model take an instruction it was never explicitly trained on. MiniMax frames language as the generalization bridge across modalities, and says it abandoned the Hailuo 02 architecture despite its advantages because task generalization is "an irreversible trend" — the same bet that moved [large language models](https://howaiworks.ai/glossary/large-language-model) from per-task fine-tunes to general instruction following. ## The Technical Choices Three changes carry the claimed efficiency gains, all reported by MiniMax without an accompanying technical report or third-party benchmark: - **H3-VAE.** A rebuilt [tokenizer](https://howaiworks.ai/glossary/tokenization) that MiniMax says delivers a 4× gain in effective sequence length at improved reconstruction quality. This is what makes native 2K affordable — fewer tokens per frame is the whole ballgame in video [autoencoders](https://howaiworks.ai/glossary/autoencoder). - **H3-Omni Transformer.** Separating understanding from generation workloads so each gets appropriate hardware treatment. Multimodal context tripled the variance in sequence length across training samples; balancing that heterogeneous compute yielded a claimed ~30% training throughput improvement. - **In-context regeneration.** Instead of a conventional super-resolution module, the base model regenerates its own low-resolution output at 2K, reusing the original multimodal context. MiniMax argues this recovers small text and fine detail that an upscaler, which never sees the reference material, cannot reconstruct. The captioning pipeline is a detail worth noting: MiniMax says it describes relationships *between* context elements and the target video, distilling roughly 100K tokens of inference down to about 4K tokens on average. ## Pricing and Availability MiniMax's pricing claim is a comparison, not a rate card: 2K output for less than a third of what mainstream rival models charge, and 768p for under half the price of competitors' 720p. No competitor is named and no per-second figure appears in the announcement, so treat it as a directional claim until a published price list exists. H3 is reachable through hosted endpoints and third-party inference platforms; local deployment waits on the weights. ## What to Watch Two things will decide whether H3 is a landmark or a good demo. The first is whether the weights actually ship, and under what license — "subject to applicable laws and regulations" is doing real work in that sentence, and a restrictive license on a 2K video model is a different product than an Apache-2.0 one. The second is independent evaluation: the announcement contains no benchmark table, no comparison against [Seedance](https://howaiworks.ai/models/seedance), [Kling](https://howaiworks.ai/models/kling) or [Wan](https://howaiworks.ai/models/wan), and no technical report. Every efficiency and quality number above is MiniMax's own. ## Conclusion H3 is MiniMax's argument that video generation should follow the same path language models took: one general model that reads instructions, rather than a catalog of task-specific endpoints. The unified design, the 4× tokenizer gain and the in-context 2K regeneration are a coherent set of choices toward that, and native stereo audio at 15 seconds is a real capability step over the silent 1080p of [Hailuo 2.3](https://howaiworks.ai/models/hailuo). Whether it changes anything outside MiniMax's [own platform](https://howaiworks.ai/ai-tools/minimax) depends entirely on the weights landing with a usable license. Until then, the most consequential claim in the announcement is the one that has not been delivered yet. ## Sources - [MiniMax H3: An Open Model Breaking the Boundaries Between Tasks and Modalities](https://www.minimax.io/blog/minimax-h3) — MiniMax, July 31, 2026 - [MiniMaxAI/MiniMax-H3](https://huggingface.co/MiniMaxAI/MiniMax-H3) — the open FL2VA and Ref2VA checkpoints and the MiniMax H3 Community License, published August 3, 2026 *(update)* - [MiniMax video pay-as-you-go pricing](https://platform.minimax.io/docs/guides/pricing-paygo) — the per-second H3 and H3-Max rate card *(update)* - [China's MiniMax releases H3 video model](https://www.aol.com/articles/chinas-minimax-releases-h3-video-044250000.html) — Reuters via AOL - [MiniMax unveils H3 video generation model amid open-weight push](https://www.digitaltoday.co.kr/en/view/87931/minimax-unveils-h3-video-generation-model-amid-open-weight-trend) — Digital Today ## Frequently Asked Questions ### What is MiniMax H3? H3 is a general-purpose multimodal generation model announced by MiniMax on July 31, 2026. It accepts text, images, video and audio in a single prompt and generates video up to 15 seconds at 2K resolution with native stereo sound. ### Are MiniMax H3's weights open? Yes, since August 3, 2026, with two caveats. `MiniMaxAI/MiniMax-H3` on Hugging Face carries the 33B FL2VA and Ref2VA base checkpoints under the MiniMax H3 Community License, but `H3-Context-IR` and `H3-Regenerate-2K` were withheld, so a local install caps at a 768-pixel short edge. This article, dated August 2, was written the day before the release. ### Can I run MiniMax H3 locally in the US or the EU? Not under the license as published. The MiniMax H3 Community License defines the European Union, the United Kingdom, the Republic of Korea and the United States as Excluded Territories, and separately requires written authorization from MiniMax for anyone whose products generate more than $20 million a year. The hosted API is not restricted this way. Read the license before deploying; this is a summary, not legal advice. ### How is H3 different from Hailuo 2.3? Hailuo 2.3 is a closed-weight video model that generates silent video up to 1080p. H3 replaces the separate task-specific models with one system that handles image, audio and video generation, editing and reference-based tasks, and produces audio jointly with the video. ### How much does MiniMax H3 cost? MiniMax's pay-as-you-go rate card, published after this article, bills H3 per second of output: $0.13/second at 2K ($7.80 per minute) and $0.08/second at 768P. A second variant, MiniMax-H3-Max, runs $0.08/second at 768P and $0.05/second at 480P and supports text-to-video and image-to-video only. ### What can you put into a single H3 prompt? According to reporting on the launch, a single generation can combine up to nine images, three video clips and three audio tracks alongside the text instruction, with the relationships between them described in natural language. --- Source: https://howaiworks.ai/blog/minimax-h3-open-weights-video-model — HowAIWorks.ai