AnswersPublished20 sources
MiniMax H3 Open Weights: What the H3-Base Release Actually Includes
MiniMax’s August 3, 2026 Hugging Face release made H3 Base available: a 33B dense, single stream model for 768p video with synchronized stereo audio. H3 takes mixed text, image, video, and audio context, while the full system can produce clips up to 15 seconds with native stereo audio and up to 2K output.
Published byEdited with GPT-5.6 TerraImages generated with GPT Image 2
Research answer
Key takeaways
- MiniMax’s August 3, 2026 Hugging Face release made H3 Base available: a 33B dense, single stream model for 768p video with synchronized stereo audio.
- H3 takes mixed text, image, video, and audio context, while the full system can produce clips up to 15 seconds with native stereo audio and up to 2K output.
- The open weights create room for self hosting and customization, but teams need to assess the community license terms, deployment hardware, and the missing closed components before treating H3 as a fully local 2K prod...
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What did MiniMax’s August 2026 open-weight release of H3-Base on Hugging Face include—covering H3’s 33B-parameter dense single-stream H3-Omn. Article summary: MiniMax’s August 3, 2026 Hugging Face release made the **H3-Base generation weights** available, not the entire three-stage H3 system. It brought a large multimodal, audio-native video model into local and customizable w. Topic tags: general, education, general web, user generated. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, cha
MiniMax’s August 2026 Hugging Face launch is best understood as an open-weight release of H3-Base, the generation stage of its H3 video system—not a release of every model needed to reproduce MiniMax’s complete hosted workflow. The distinction matters: H3-Base can generate synchronized video and audio at 768p, while the system’s context-processing and 2K regeneration stages remained separate. 1
5
What MiniMax released
MiniMax published H3 weights to Hugging Face on August 3, 2026. The available release centered on H3-Base, including reported FL2VA and Ref2VA task weights. 1
13
At its core is the H3-Omni-Transformer, a 33B-parameter dense, single-stream Transformer. Instead of treating audio as an afterthought, the model jointly denoises video and audio in a shared generation process. The broader H3 system accepts unified context made from text, images, video, and audio, and is designed to return video with native stereo sound. 5
10
The important caveat is that “33B” describes the Omni-Transformer generation core, not necessarily every component a complete deployment must load. The pipeline also uses an H3 encoder plus VisualVAE and AudioVAE components. 10
13
How the H3 pipeline is divided
MiniMax describes H3 as a three-part system:
- H3-Context-IR interprets and refines complex multimodal instructions into a context intermediate representation for generation. MiniMax says this stage is important to final quality.
5
- H3-Base generates the synchronized audio and video result at 768p from that processed context. This is the component made available as open weights.
13
- H3-Regenerate-2K performs the subsequent regeneration step used for 2K output. It was not part of the open-weight release.
9
That means a local H3-Base installation should not be confused with the entire H3 product shown in 2K demos. The full system supports output up to 2K and up to 15 seconds, but the released base generator is the 768p stage. 5
9
Architecture: one generative stream for sound and picture
H3’s central design is a single stream for multimodal video-and-audio generation. The 33B H3-Omni-Transformer uses a dense Transformer architecture; MiniMax’s repository says roughly 13B parameters are in AdaLN-related branches, whose outputs can be precomputed and cached. 10
The supporting representation pipeline includes an encoder, a visual VAE, and an audio VAE. Reported technical descriptions say the visual VAE compresses video by 16× spatially and 4× temporally; the tokenization design yields an effective 32× spatial reduction. These compression stages are fundamental to making video generation tractable, but they also illustrate why the headline parameter count is not a complete measure of pipeline size or runtime needs. 1
14
What H3 can do—and what the open release cannot do alone
The full H3 system is built for prompts and references that mix text, still images, video clips, and audio. It can generate clips with native stereo audio at resolutions up to 2K and durations up to 15 seconds. 5
10
For builders, the open H3-Base weights enable experimentation with local inference, quantization, adapters, fine-tuning approaches, and custom creative tooling. But the released model does not include MiniMax’s Context-IR system or its 2K regeneration model. Teams using complicated references may therefore need their own prompt/context preparation layer, and teams requiring 2K output cannot assume that H3-Base alone reproduces the hosted result. 5
9
Pricing and local deployment claims
Published coverage reported a 2K API price of 0.8 yuan per second and stated that audio references and the first five image references were free under the referenced offering. 14
27 Pricing is service- and region-dependent, so these figures should be treated as a published point-in-time API rate rather than a permanent global price list.
MiniMax and SGLang promotional material also characterized H3 as costing roughly one-third as much as Seedance 2.0 in a particular comparison, and said it could run locally on two RTX 5090 GPUs or one RTX 6000. 28
31 Those are deployment claims for a particular software, precision, and offloading setup—not universal minimum hardware requirements. The total workload includes the encoder and VAE components in addition to the 33B transformer.
13
Why the release matters for production teams
More control over sensitive inputs
A self-hosted workflow can keep product imagery, unreleased footage, voice assets, and prompts within an organization rather than sending them to a third-party generation API. That can be attractive for teams working with confidential assets. It does not remove the need for access controls, input-rights review, output review, and internal governance.
A starting point for custom pipelines
Open weights give technical teams a path to connect generation with asset-management libraries, creative-review systems, rendering tools, and post-production workflows. They can also investigate quantization and task-specific adaptation. However, organizations should review the license before committing to a deployment: third-party reporting describes geographic restrictions affecting local use in the US, EU, UK, and South Korea. 4
Strongest fit: iterative, short-form creative work
The practical fit is clearest for teams making many short variations: advertising creatives, e-commerce product clips, concept work, animatics, and social-content experiments. A 768p local generator with integrated audio can reduce friction in early creative iteration, particularly where teams already have suitable GPU capacity.
It is less of a turnkey replacement for every video-production requirement. The open release lacks the full context and 2K stages, and a short-clip model still benefits from human review, editing, and rights management. The real value is not simply that H3 is downloadable; it is that a capable audio-video generation stage can be incorporated into a controlled workflow—where the license and hardware allow it.
Bottom line
MiniMax opened a meaningful part of H3: H3-Base, powered by a 33B single-stream Transformer that jointly generates video and native stereo audio from multimodal inputs. But the release was deliberately narrower than the full H3 product. Context-IR and 2K regeneration remained closed, making H3-Base a customizable 768p generation foundation rather than a complete open 2K system. 5
9
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "MiniMax H3 Open Weights: What the H3-Base Release Actually Includes"?
MiniMax’s August 3, 2026 Hugging Face release made H3 Base available: a 33B dense, single stream model for 768p video with synchronized stereo audio.
What are the key points to validate first?
MiniMax’s August 3, 2026 Hugging Face release made H3 Base available: a 33B dense, single stream model for 768p video with synchronized stereo audio. H3 takes mixed text, image, video, and audio context, while the full system can produce clips up to 15 seconds with native stereo audio and up to 2K output.
What should I do next in practice?
The open weights create room for self hosting and customization, but teams need to assess the community license terms, deployment hardware, and the missing closed components before treating H3 as a fully local 2K prod...