Summary: MiniMax announced H3 on July 31 as an open 33B omni-modal video model supporting text-to-video, image-to-video, first/last-frame control, references, and native stereo audio. Clips reach 15 seconds at 24 fps, defaulting to 768p and rising to 2K through hosted regeneration. Weights reached Hugging Face on August 3, but only H3-Base is open; instruction parsing and 2K regeneration remain hosted.
MiniMax introduced H3 on July 31, billing it as the successor to Hailuo 01 and 02 and launching it under the tagline “An Open Model Breaking the Boundaries Between Tasks and Modalities.” Calling it just another video model would miss the point: text, images, video, and audio all enter one shared context, and what comes back out is video carrying native stereo sound. The decisive sentence in the announcement was a promise to publish the weights “in the coming days” — and on August 3 they landed on Hugging Face, with ComfyUI shipping Day-0 support that same day.
For commercial video models, the dominant pattern remains closed weights: you call an API and pay by the second. H3 took “open source” from a stated vision and turned it into a schedule with dates attached — then met it.
From Hailuo to H3: Betting on Task Generalization
The announcement states the central claim plainly: “We believe task generalization is an irreversible trend, and architectural tricks should give way to how the model is defined.” Reduced to features, that means one model covers text-to-video, image-to-video, first-and-last-frame conditioning, and reference-to-video, plus spoken dialogue, sound effects, and music — produced in a single pass, so no separate TTS, music, and SFX stitching happens afterward.
The specs bear that out: up to 15 seconds per clip at 24 fps with 32 kHz stereo audio, 768p by default and up to native 2K. How 2K is reached deserves a closer look. Rather than attach a super-resolution module, H3 has its base model regenerate the low-resolution output it has just produced — In-Context Regeneration, in MiniMax’s naming — and the four-times effective sequence length gain from H3-VAE is what makes that workable. On pricing, MiniMax’s claim is that a second of 2K output costs less than a third of what mainstream models charge.
Inside the 33B Omni-Transformer
Once the weights were up, the technical details surfaced on the Hugging Face model card and on GitHub. The core is a dense, single-stream Omni-Transformer of 33B parameters, with roughly 13B of that allocated to AdaLN-related branches that can be cached or skipped during inference. The encoder reuses the full Qwen3-VL-32B weights and draws hidden states from its 50th layer — an open model stacked on top of another open model, in a form you can point at.
Position information comes from 3D Multimodal RoPE, spanning time, height, and width. A VisualVAE compresses 16x spatially and 4x temporally; an AudioVAE converts 32 kHz stereo into 40 Hz latent tokens, so audio shares a sequence with the picture and is trained and generated alongside it. Two open checkpoints map onto two workflows: FL2VA takes text or first/last frames, while Ref2VA takes up to 9 images, 3 video clips, and 3 audio clips as an omni-reference. Dialogue is supported in 11 languages.
Open, But Not Everything
H3 ships under MiniMax’s Community License, and what counts as open has to be stated precisely: of the three-module architecture, only H3-Base is open. H3-Context-IR, which parses instructions and handles preprocessing, runs as a hosted service. H3-Regenerate-2K, the module that lifts output to 2K, is not open yet, and the trained sparse attention has been deferred. The complete 768p pipeline is what you actually get; 2K still goes through the official API.
This “open trunk, hosted key modules” split is the pragmatic form the open video model takes in 2026. Read one way, it looks like good faith at a discount; read the other way, the community still ends up with a body it can fine-tune and wire into its own pipelines, rather than a web form.
Running on Consumer GPUs
H3’s explosion in the community came from ComfyUI’s Day-0 support, shipped the day the weights dropped. The slimming recipe in the official post is blunt about the mechanics: about 40% of the modulation weights were pruned and replaced with a functionally equivalent lookup table; int8 quantization plus custom kernels then cut peak VRAM from 123.6 GB at full precision to 42.5 GB — a 66% reduction. Add dynamic offloading and a single RTX 3060 can run it locally; on the SGLang Diffusion side, the recipe is two RTX 5090s or one RTX Pro 6000.
A single line from the ComfyUI post sums up what the model changes: “Audio is a property of the model, not a post-process.” Because video and sound are born inside one model, sync is guaranteed by construction.
What It Means for Creators and Developers
Three things are worth noting. First, generating audio and video jointly shortens the post-production chain — dialogue, sound effects, and music no longer need separate generation passes followed by alignment, and dialogue in 11 languages is built in. Second, an open video model that runs on consumer GPUs makes local pipelines, offline scenarios, and cost-controlled batch workflows realistic for the first time, and the reuse of a Qwen3-VL encoder is a reminder that the open ecosystem compounds. Third, when evaluating an “open model,” read the license and the module boundaries: H3 opens the trunk while keeping 2K and instruction parsing on the official side, and that split directly shapes how you design your architecture. MiniMax has also said the next H generation will integrate capabilities from its M-series LLMs — the fusion of video models and LLMs is just getting started.