On July 31, MiniMax announced H3, the successor to Hailuo 01 and 02, under the tagline “An Open Model Breaking the Boundaries Between Tasks and Modalities.” This is not just another video model: text, images, video, and audio go into one shared context, and the output is video with native stereo sound. The crucial part of the announcement was a commitment to release the weights “in the coming days” — and on August 3 they landed on Hugging Face, with ComfyUI shipping Day-0 support the same day.
The mainstream route for commercial video models is closed weights: you call an API and pay by the second. H3 turned “open source” from a vision into a schedule, and then kept it.
From Hailuo to H3: Betting on Task Generalization
The core claim sits in the announcement: “We believe task generalization is an irreversible trend, and architectural tricks should give way to how the model is defined.” In practice, 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 — all generated in one pass, with no separate TTS, music, and SFX stitching afterward.
The specs hold up: up to 15 seconds per clip at 24 fps with 32 kHz stereo audio, 768p by default and up to native 2K. The 2K approach is worth noting: instead of bolting on a super-resolution module, H3’s base model regenerates its own freshly produced low-resolution output — MiniMax calls it In-Context Regeneration. Combined with the four-times effective sequence length gain from H3-VAE, native 2K becomes feasible. On pricing, MiniMax claims 2K output costs less than a third of mainstream models per second.
Inside the 33B Omni-Transformer
Once the weights went up, the technical details landed on the Hugging Face model card and GitHub. The core is a 33B-parameter dense, single-stream Omni-Transformer, with roughly 13B of that sitting in AdaLN-related branches that can be cached or skipped at inference. The encoder reuses the full Qwen3-VL-32B weights, feeding hidden states from its 50th layer — one open model stacked on another, in a very concrete form.
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 sound lives in the same sequence as the picture and is trained and generated jointly. The two open checkpoints map to 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 the open portion needs to be stated precisely: of the three-module architecture, only H3-Base is open. H3-Context-IR, which parses instructions and handles preprocessing, is 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. What you actually get is the complete 768p pipeline; 2K still goes through the official API.
This “open trunk, hosted key modules” pattern is the pragmatic version of the open video model in 2026. Critics can call it good faith at a discount; from the other side, the community at least gets a body it can fine-tune and wire into its own pipelines, rather than a web form.
Running on Consumer GPUs
What made H3 explode in the community was ComfyUI’s Day-0 support on the day the weights dropped. The slimming recipe described in the official post is blunt: about 40% of the modulation weights were pruned and replaced with a functionally equivalent lookup table; int8 quantization plus custom kernels cut peak VRAM from 123.6 GB at full precision to 42.5 GB — 66% less. With dynamic offloading, a single RTX 3060 can run it locally; on the SGLang Diffusion side, the recipe is two RTX 5090s or one RTX Pro 6000.
One line from the ComfyUI post sums the model up: “Audio is a property of the model, not a post-process.” Video and sound are born in the same model, so sync is guaranteed by construction.
What It Means for Creators and Developers
Three things to note. First, joint audio-video generation shortens the post-production chain — dialogue, sound effects, and music no longer need to be generated separately and aligned, 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; 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 but keeps 2K and instruction parsing on the official side, and that directly shapes 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.
Sources
- MiniMax H3: An Open Model Breaking the Boundaries Between Tasks and Modalities — MiniMax
- MiniMax-H3 model card — Hugging Face
- MiniMax H3 Day-0 Support in ComfyUI — ComfyUI Blog
- MiniMax-AI/MiniMax-H3 — GitHub
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.