MiniMax-H3 Audio-Visual Video Generation Model: 32GB+ VRAM, File List

September 26, 2026September 27, 2026

Sample outputs are available on the model card.

At a Glance

Values determined by this site’s code at collection time. Dates are JST.

Overview

MiniMax has released the weights for MiniMax H3 on Hugging Face, a model that accepts text, image, video, and audio as inputs and generates videos with audio. The model card describes H3 as a “general-purpose omnimodal generation system capable of producing videos up to 15 seconds long with a maximum resolution of 2K, featuring native stereo audio." Capabilities include text-to-video generation, specifying start and end frame images, and providing reference images, videos, and audio.

However, only the core generation model, H3-Base, of the overall system has been released. H3-Context-IR, which interprets and structures input prompts, and H3-Regenerate-2K, which upscales the output to 2K, are not included in this release and are provided via the MiniMax API. Running locally is limited to 768p generation using H3-Base.

Specifications

• Generation model (H3-Omni-Transformer): A 33B parameter dense single-stream Transformer. Approximately 13B of these parameters are in the AdaLN-related branches. According to the model card, AdaLN outputs can be precomputed and cached, meaning the weights for this portion do not need to be loaded for inference-only use (all weights are released for purposes such as fine-tuning).
• Architecture: Attention and FFN layers have no modality-specific structures; modality-specific parameters exist only in the input/output layers and AdaLN branches. Positional representation uses 3D MM-RoPE (time, height, width).
• Text encoding (H3-Encoder): Uses the pretrained weights of Qwen3-VL-32B wholesale, passing the hidden states of its 50th layer to the generation model. Because special tokens for H3 are added, it is necessary to use the tokenizer and configuration files included in the H3 repository.
• Video VAE (H3-VisualVAE): A video autoencoder that is causal in the temporal dimension, compressing spatially by 16x and temporally by 4x, with 24 latent channels (f16t4d24). It further groups them into 1 × 2 × 2 patches, resulting in a spatial reduction of 32x and a temporal reduction of 4x when entering the Transformer. The decoder is ViT-based and trained separately.
• Audio VAE (H3-AudioVAE): Handles stereo by processing the left and right channels separately using the same encoder and decoder. Compresses 32 kHz audio into sequences of 40 Hz latents per channel.
• Output specifications: Length is 4 to 15 seconds, frame rate is 24 FPS, audio is 32 kHz stereo. Aspect ratios supported include 21:9, 16:9, 4:3, 1:1, 3:4, 9:16, etc., with a default resolution of 768 pixels on the short side (2K uses H3-Regenerate-2K).
• Dialog languages: Reportedly supports 11 languages stably (Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, Spanish), with other languages supported to some extent.
• Released weights: Two checkpoints categorized by use case, both in BF16. These are weights distilled from CFG.
• H3-Base-FL2VA: Text-to-audio-video (t2va) and generation from start and end frames (fl2va). Takes 0 to 2 input images: 0 for text-only, 1 for start or end frame, 2 for both boundary frames. – H3-Base-Ref2VA: Reference-based generation (ref2va). Accepts up to 9 images, 3 videos (2 to 15 seconds each, up to 15 seconds total), and 3 audio files (2 to 15 seconds each, up to 15 seconds total), for a total of up to 12 files.
• Sparse attention: Introduced in the final stage of training, but the initial release only supports full-attention for inference. Implementations for sparse attention are planned for release at a later date.
• License terms: Distributed under the “MiniMax H3 Community License." The model card notes that this license includes obligations regarding lawful use and usage restrictions. Because specific terms such as commercial use availability are not written in the body of the model card, users should review the original license text before use.

Performance

Benchmarks or human evaluation tables comparing it against other models are not included in the model card. Instead, it provides reproducible examples where the same requests (three examples covering T2VA, FL2VA, and Ref2VA) were generated using both local H3-Base and the MiniMax API for side-by-side comparison. Each example lists the prompt structured by H3-Context-IR, the 768p output from H3-Base, the 2K output from H3-Regenerate-2K, and reference outputs generated directly via the API at 2K and 768p.

Regarding quality, the model card states: “Thanks to a design that emphasizes generalizing tasks, it can understand and generate a wide range of multimodal contexts starting from the pretraining stage, demonstrating high performance in following complex multimodal instructions." Since no quantitative backing is provided, this should be taken as a claim from the publisher.

Additionally, the model card emphasizes that H3-Context-IR heavily dictates final output quality. It is the process of interpreting the relationships among input images, audio, and reference videos alongside the flow of time, filling in missing descriptions, and converting them into structured instructions that H3-Base can easily understand. The prompt examples are meticulously detailed down to shot allocation, sound effects, and musical scores. Short prompts passed directly to local H3-Base should not be expected to match API-level quality.

Strengths and Use Cases

Tags such as text-to-video, image-to-video, video-to-video, text-to-audio-video, reference-to-audio-video, and synchronized-audio-video line up, highlighting simultaneous video and audio generation as its primary feature. Sound effects, environmental audio, musical scores, and dialogue can be generated in sync with the video. For instance, an example specifies engine hums, impact noises, and an orchestral score for a scene where a fleet goes to warp on a spaceship bridge. Because it stably supports dialogue in 11 languages, it can also be used for short videos containing Japanese dialogue.

Selection is determined by two checkpoints:

• FL2VA is used when defining start and end frames like a storyboard to bridge the gap, or when creating solely from text.
• Ref2VA is used when building from multiple clues, such as character images, voice or music references, or continuing an existing video.

For finishing in 2K, a “Full 2K Workflow" is suggested, where local H3-Base 768p outputs are passed to H3-Regenerate-2K via the MiniMax API for redrawing. In this case, prompt interpretation (H3-Context-IR) and 2K upscaling are handled on the MiniMax side, subject to automated content moderation (blocking illegal, explicit, or potentially infringing content).

How It Differs from Similar Models

Previously, our site covered “Minimax-h3_Singularity”. That was a fine-tune where an independent developer condensed text, image, reference, and video generation into a single model running in ComfyUI, incorporating pruning and weight adjustments. This article covers the official MiniMax weights that served as its base, which are split into two checkpoints by use case.

Hardware Requirements

Estimated requirements (calculated by Local Model Watch)

Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. File sizes are measured from the converted build unsloth/MiniMax-H3-GGUF. For this generative model, the file size is the total of the model itself plus the smallest text encoder in the repository (qwen3vl_32b_minimax_h3-Q2_K_M, 12.2GB) and the VAE (minimax_h3_audio_vae_fp32 (0.6GB), minimax_h3_video_vae_fp16 (4.9GB)), assuming all parts are loaded at once. Tools that load the parts one at a time, such as ComfyUI, can get by with less. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.

Can You Run It Locally?

The publisher distributes this model as safetensors.

License — other: A custom license from the publisher. Check the original terms directly, including whether commercial use is permitted.

Compiled by this site’s code from the published formats and the license field. License summaries are not legal advice — check the publisher’s original terms before relying on them.

Distributed Files

Weight files published in MiniMaxAI/MiniMax-H3, listed by this site from the Hugging Face API. Sizes are the actual file sizes.

How to Get It

• Distribution format: Original-format checkpoints (FL2VA/ and Ref2VA/) alongside diffusers-format files are hosted at Hugging Face’s MiniMaxAI/MiniMax-H3. Each checkpoint contains the generation model along with the processor, tokenizer, text encoder, visual VAE, and audio VAE. Downloading does not require agreeing to terms of service.
• Download command (when taking the original format for SGLang/vLLM):
• Both: hf download MiniMaxAI/MiniMax-H3 --include "model_index.json" "FL2VA/*" "Ref2VA/*" --local-dir MiniMax-H3 – Single checkpoint: hf download MiniMaxAI/MiniMax-H3 --include "model_index.json" "FL2VA/*" --local-dir MiniMax-H3
• When using diffusers, manual downloading is unnecessary; ModularPipeline.from_pretrained("MiniMaxAI/MiniMax-H3") fetches only the required components.
• Supported tools: SGLang, vLLM, diffusers, and ComfyUI are recommended by the model card. ComfyUI features workflow templates for reference-based generation (R2V) and text-to-video generation (T2V).
• SGLang startup example: sglang serve --model-path MiniMaxAI/MiniMax-H3 --num-gpus 4 --ulysses-degree 4 --performance-mode speed --model-variant fl2va (use --model-variant ref2va for Ref2VA). The model card example assumes parallel execution across 4 GPUs.
• Prompt writing: For users structuring prompts themselves without H3-Context-IR, the model card publishes a guide on writing basic and reference prompts, along with a collection of skills to improve prompts.

Quantized and Converted Variants

File sizes of each build:

• Available builds in unsloth/MiniMax-H3-GGUF: Q2_K 6.2GB / Q2_K_XL 7.5GB / Q3_K 8.1GB / Q3_K_XL 8.9GB / Q4_K 10.6GB / Q2_K_M 12.2GB / Q5_0 12.9GB / Q6_K 15.4GB / Q4_K_M 17.0GB / Q8_0 19.9GB

In addition, 49 converted build(s) from other uploaders exist on Hugging Face; this site lists only builds from the model’s publisher or established quantization maintainers.

This section is appended automatically by Local Model Watch when a converted build of this model appears after publication. Memory figures are estimated from the size of the distributed files. What the quantization names mean: glossary.

Other Models for the Same Task

Recent video generation models covered by Local Model Watch, newest first. Grouped by the task each publisher declares on Hugging Face (pipeline_tag); the smallest VRAM tier is this site’s estimate.

• Lightricks Releases Multiple LoRA and IC-LoRA Adapters for LTX-2.5

See all video generation models →

What to Read Next

• Find models by VRAM (This model needs at least 32GB) → VRAM quick reference
• Explore the same model family → MiniMax-H3 family overview (2 articles, 1 converted builds)
• What Q8_0 mean and where to get this model → GGUF format guide and models / Quantization and model-format glossary
• Learn about the publisher → MiniMax: models, licenses and articles

Sources

• https://huggingface.co/MiniMaxAI/MiniMax-H3

Update History

• 2026-09-26: Added converted builds to “Quantized and Converted Variants”: unsloth/MiniMax-H3-GGUF
• 2026-09-26: Updated the hardware requirements table with the actual file sizes of unsloth/MiniMax-H3-GGUF.
• 2026-09-26: Changed the title to show what the article covers (VRAM requirements, file list, etc.).