MiniMax released H3, its third-generation video model, with open weights on August 3, 2026. ComfyUI, the popular node-based image and video generation interface, announced day-zero support starting in version 0.30.0. The model accepts text, images, video, or audio as input and generates video with native stereo sound at up to 2K resolution and up to 15 seconds per clip. H3 is the first MiniMax video model released with open weights, succeeding Hailuo 01 and Hailuo 02 in the company's video generation lineage (ComfyUI blog).

MiniMax officially launched H3 on July 31, describing it as a general-purpose omni-modal generation model — meaning it can understand and work across multiple input types, including language, video, speech, and music, all within a single model (MiniMax blog). The company's homepage characterizes H3 as open-weight and general-purpose, and its API platform lists it as a new-generation open general-purpose multimodal video model.

The model supports several generation workflows: text-to-video, image-to-video, first-and-last-frame (where you provide a starting and ending image and the model fills in the motion between them), and reference-to-video (where a sample input guides the style or content of the output). ComfyUI provides downloadable workflows for image-to-video, reference-to-video, and text-to-video out of the gate. H3 generates audio natively in the same forward pass as video, in stereo, rather than as a separate post-processing step. This is a structural design choice, not a pipeline convenience, and it sets H3 apart from the common approach of attaching a separate audio model to a video model after the fact.

ComfyUI's engineering team applied two categories of optimization to bring H3's inference (the process of running the model to produce output) within reach of consumer hardware. First, they pruned H3's modulation weights, which account for roughly 40% of the model's total parameters. Parameters are the learned values a neural network uses to make decisions; modulation weights are a specific subset that adjusts how the model responds to different inputs. The team replaced these with a functionally equivalent lookup table — a precomputed reference that produces the same results without storing all those learned values in memory. Second, the shipped weights include int8 convrot quantization paired with custom kernels. Quantization reduces the precision of the numbers the model uses (from higher-bit floating point down to 8-bit integers), shrinking memory requirements, while the custom kernels (specialized pieces of code) reduce peak VRAM use during inference (ComfyUI blog).

Combined, these optimizations reduce H3's total memory footprint by 66%, from 123.6 GB in full precision down to 42.5 GB with the smallest model variants. With dynamic VRAM offloading — a technique that moves parts of the model in and out of GPU memory as needed — ComfyUI reports that H3 can run locally on a GPU as modest as an RTX 3060. That is a meaningful threshold: it places a model capable of 2K video with synchronized stereo audio within reach of a large installed base of consumer GPUs, not just data-center hardware.

Model weights are available in two repositories on Hugging Face, a popular platform for sharing machine learning models. MiniMax's original repository is hosted at MiniMaxAI/MiniMax-H3. ComfyUI hosts a repackaged version at Comfy-Org/MiniMax-H3, which includes the pruned modulation weights, int8 quantized convrot kernels, and ComfyUI-compatible model files. On Hugging Face, H3 is categorized as an Image-Text-to-Video model.

The broader context here is where generative video has been heading. Over the past two cycles of video model releases, the frontier has been defined by progressively higher resolution and longer clip durations, but access has been gated behind API-only delivery or weights requiring multi-GPU inference. H3's open-weights release, paired with ComfyUI's 66% memory reduction, shifts that gate. A developer with a single consumer-tier GPU can now download weights, run inference locally, and chain H3 into node-based pipelines alongside other models, all without an API key.

The native audio generation in the same pass as video has practical implications for how pipelines are built. When audio is generated jointly rather than after the fact, temporal synchronization is handled at the model level, eliminating the need for a separate audio model, an alignment step, or heuristic-based lip-sync or sound-effect matching. For ComfyUI users building multi-node workflows, this reduces graph complexity and removes a class of failure modes around audio-video desynchronization.

The omni-modal input scope also broadens what a single model can do within a pipeline. Accepting text, images, video, or audio as input means H3 can serve as a conditioning stage for downstream models or as a terminal generation step, depending on workflow design. Reference-to-video, in particular, enables style and content transfer from a reference input without a separate embedding or encoder step external to the model.

ComfyUI's day-0 support track record has become a signal in the open-weights generative model ecosystem. When a model ships with ComfyUI integration on release day, it immediately enters a distribution channel that reaches a substantial community of developers, artists, and researchers who build and share node-based pipelines. The availability of pre-built downloadable workflows for three of H3's generation modes lowers the barrier to adoption further. MiniMax's decision to release H3 with open weights, rather than API-only as has been common for frontier video models, aligns with a pattern that has accelerated across the generative model landscape over the past year, where open-weights releases increasingly compete with proprietary offerings on capability rather than access alone.

The technical specifics of ComfyUI's optimization work merit attention from anyone building inference infrastructure for large video models. The modulation-weight pruning via lookup table replacement is not a generic quantization pass; it is a model-specific intervention targeting a structural component of H3's architecture. The int8 convrot quantization with custom kernels similarly targets a specific layer type rather than applying a blanket precision reduction. Both choices reflect an approach to memory reduction that prioritizes preserving generation fidelity at the level of individual architectural components rather than accepting a uniform quality tradeoff.

For practitioners evaluating H3, the relevant comparison points are no longer just API endpoint quality and pricing. They now include local inference feasibility, pipeline integration depth, and the degree to which a model can be modified, quantized, and composed with other open models. The combination of H3's capabilities and ComfyUI's optimizations makes that evaluation practical on hardware that many developers already own.