MiniMax H3 brings 33B open-weight video generation

MiniMax H3 is now available as an open-weight multimodal video model. The model accepts text, images, video and audio, then generates video with native stereo sound; the published checkpoints are available for local deployment. The practical split is simple: try H3 in the Hailuo AI web app first, or download the weights from the MiniMax H3 model card when you need local control.

Definition: MiniMax H3 is a 33B-parameter omni-modal generation system for video and audio.

Example: H3 can use text together with reference images, video and audio to generate a new clip or edit an existing multimodal context.

Key takeaway: H3 is not only a text-to-video checkpoint; its design unifies several reference, editing and generation tasks.

Business impact: Teams can test the model as a hosted creative tool and then evaluate a local deployment when data control or custom workflows matters.

What does MiniMax H3 generate?

MiniMax H3 generates video with native stereo audio for clips from 4 to 15 seconds, with support for multiple aspect ratios and output resolutions. The model card documents 24 FPS video and 32 kHz stereo audio, so the useful takeaway is to evaluate MiniMax H3 as an audiovisual generation system rather than as a silent animation model. Background: MiniMax Music 3 Generates Five-Minute Songs Locally.

The model's input modes are broader than a single prompt. The H3-Base FL2VA checkpoint supports text-to-video and first/last-frame generation; the Ref2VA checkpoint accepts reference images, video and audio. The model card allows up to nine images, three video clips and three audio clips in the reference workflow, with a maximum of 12 files across mixed inputs. That makes H3 relevant when a shot depends on relationships between several source assets rather than on a description alone.

Why is H3 different from a typical video model?

MiniMax H3 treats language as the bridge between modalities and tasks. In its launch explanation, MiniMax says the training design includes text-to-image, text-to-video, jointly generated audio, generalized reference and editing, and multi-shot modeling. In practical terms, a user can describe how a reference image, motion source and audio track should relate to the target clip instead of selecting a narrow task-specific mode for every operation.

This does not mean every part of the official pipeline is open. The released H3-Base checkpoints cover local 768p generation, while H3-Context-IR preprocesses complex multimodal instructions and H3-Regenerate-2K creates the higher-resolution result. MiniMax says Context-IR and Regenerate-2K are hosted or not yet open-sourced, so the complete official 2K workflow still combines local H3-Base inference with MiniMax APIs.

How can you run MiniMax H3 locally?

MiniMax publishes two task-specific H3-Base checkpoints with the processor, tokenizer, text encoder, transformer, visual VAE and audio VAE components needed for inference. The model card lists SGLang, vLLM, diffusers and ComfyUI as supported deployment paths, which gives developers several ways to trade setup simplicity for control.

The documented Hugging Face download path is:

hf download MiniMaxAI/MiniMax-H3 \
  --include "model_index.json" "FL2VA/*" "Ref2VA/*" \
  --local-dir MiniMax-H3

For a first local test, choose one task family instead of downloading both checkpoints, then follow the runtime-specific recipe. How to run an LLM locally explains the same general separation between model files, runtimes, memory and workload; H3 adds video and audio components that make the memory and latency budget much larger.

What hardware does local H3 need?

The MiniMax H3 model card does not give a single universal VRAM or unified-memory threshold. It lists a 33B-parameter model in BF16 and separate visual and audio components, so local sizing must follow the selected checkpoint and runtime rather than a headline memory number. The practical action is to test one task family at the target resolution and measure latency before committing to a larger deployment.

The memory requirement matters because the Hugging Face page lists H3 at 33B parameters and the original checkpoints in BF16. A local deployment must hold more than the nominal weights: it also needs the video and audio VAEs, runtime buffers, intermediate activations and the working inputs. The Yowox Local LLM Hardware Calculator is designed for language-model sizing rather than video diffusion, but its memory-budget principle still applies: capacity alone does not predict a usable speed.

ComfyUI’s H3 support provides a lower-friction route for experimentation. Its H3 workflows include text-to-video, image-to-video, first-and-last-frame and reference-to-video generation, while the implementation uses quantization and VRAM offloading to reduce the local footprint. That is useful for testing, but it should not be confused with the unmodified BF16 checkpoint described in the original model card.

What should MiniMax H3 creators test first?

MiniMax H3 creators should start with a short prompt and a single reference image, then compare the result with a local 768p run before attempting the full 2K pipeline. The model card separates local H3-Base output from the complete 2K workflow, so test instruction following, reference preservation, motion continuity and audio-video synchronization separately before expanding the pipeline.

MiniMax H3 is most significant as an open-weight starting point for an audiovisual workflow. It gives developers local H3-Base checkpoints, multiple inference runtimes and a path toward custom pipelines, while keeping the highest-level context processing and 2K regeneration partly hosted. That boundary is the fact to watch as the ecosystem around the model matures.