MiniMax-H3 Guide: VRAM Requirements, GGUF Builds, Model Files
About This Model
MiniMax H3 is MiniMax’s model for generating video with sound. From instructions that combine text, images, video and audio, it creates videos of up to 15 seconds with native stereo audio. Its defining feature is that sound effects, ambience, music and dialogue are generated together with the picture, and dialogue is reliably supported in 11 languages, including Arabic, Chinese, English and Japanese.
The full system has three parts: H3-Context-IR, which interprets and refines the instructions; H3-Base, which generates the video and audio at 768p; and H3-Regenerate-2K, which redraws the result at 2K. Only H3-Base has been released as open weights.
At the core of H3-Base is a 33B-parameter dense Transformer (H3-Omni-Transformer), and text is encoded with the full pretrained weights of Qwen3-VL-32B. The weights come as two task-specific checkpoints:
- FL2VA: generates from text alone, or from first and/or last frame images (0–2 images).
- Ref2VA: generates from references: up to 9 images, 3 videos and 3 audio clips.
What Makes It Stand Out
- One model makes the picture and the sound together. A single-stream Transformer predicts video and audio latents jointly, with 32 kHz stereo audio. There is no need to add sound afterward with a separate model.
- It is flexible about references. Ref2VA accepts up to 12 files in total across images, video and audio, so it can work from a character’s appearance, a voice, or footage to continue.
- The output specs are practical. Clips run 4–15 seconds at 24 FPS, and aspect ratios from 21:9 to 9:16 are supported.
- It can be fine-tuned. All weights are released, including the AdaLN branches (about 13B parameters) that inference-only deployments can skip.
Caveats and Weaknesses
- There are no numbers to back up its quality. The model card has no benchmark or human-evaluation comparison with other models, so the claim that it “follows complex multimodal instructions" well is the publisher’s own.
- Local quality is not the same as the API’s. The model card stresses that H3-Context-IR, which refines the instructions, strongly shapes the final quality. That part is not released: you either use it through the API or refine instructions yourself following the published prompting guides. The 2K redraw is also API-only, so local output tops out at 768p.
- Sparse attention is not available yet. It was used in training, but the initial release supports inference with full attention only, so longer videos cost more compute.
Running It Locally
- You need a large GPU, or a tool that loads the parts one at a time. Besides the model itself, it needs a text encoder (Qwen3-VL-32B) plus video and audio VAEs. This site’s memory table (further down this page) is the total for loading the model, the smallest text encoder and the VAEs from the GGUF build all at once. Tools that load the parts one at a time, such as ComfyUI, can get by with less.
- The publisher recommends SGLang, vLLM, diffusers and ComfyUI. ComfyUI has workflow templates for text-to-video (T2V) and reference-to-video (R2V). The SGLang example runs in parallel on four GPUs.
- Detailed instructions that spell out shots, sound and music get closer to the publisher’s examples than short prompts do.
- The license is the “MiniMax H3 Community License." The model card mentions obligations on lawful use and use restrictions, but it does not spell out concrete terms such as whether commercial use is allowed. Check the license text before use.
- This family also includes a fine-tune packaged for ComfyUI by an individual developer (Minimax-h3_Singularity). Its weights differ from the official model; it is a separate model.
Source: MiniMaxAI/MiniMax-H3 model card, as of 2026-09-26.
Our Coverage and Data
Everything Local Model Watch has published about the MiniMax-H3 family: 2 article(s) covering the base model and its fine-tunes, plus converted builds we tracked after publication. Memory requirements below are computed by this site from file sizes, not quoted from model cards. Part of our model family index.
At a Glance
| Item | Value |
|---|---|
| Base model(s) | MiniMaxAI/MiniMax-H3 |
| Publisher | MiniMax |
| License (model card) | apache-2.0, other |
| Smallest VRAM tier | 32GB |
| Articles | 2 |
Hardware Requirements
Estimated requirements (calculated by Local Model Watch)
| Your VRAM | Quantization | File size | Est. memory needed |
|---|---|---|---|
| 32GB (RTX 5090, etc.) | Q3_K_XL | 26.5GB | 31.8GB |
| 48GB (RTX 6000 Ada / A6000, etc.) | Q8_0 | 37.6GB | 45.1GB |
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.
| File | Size |
|---|---|
audio_vae/diffusion_pytorch_model.safetensors |
605MB |
FL2VA/audio_vae/model.safetensors |
605MB |
FL2VA/text_encoder/model-*-of-00014.safetensors |
66.71GB (14 split files) |
FL2VA/transformer/model-*-of-00013.safetensors |
66.28GB (13 split files) |
FL2VA/video_vae/source/model.safetensors |
10.42GB |
Ref2VA/audio_vae/model.safetensors |
605MB |
Ref2VA/text_encoder/model-*-of-00014.safetensors |
66.71GB (14 split files) |
Ref2VA/transformer/model-*-of-00013.safetensors |
66.28GB (13 split files) |
Ref2VA/video_vae/source/model.safetensors |
10.42GB |
text_encoder/model-*-of-00014.safetensors |
66.71GB (14 split files) |
transformer/diffusion_pytorch_model-*-of-00014.safetensors |
66.28GB (14 split files) |
transformer_ref/diffusion_pytorch_model-*-of-00014.safetensors |
66.28GB (14 split files) |
vae/diffusion_pytorch_model-*-of-00003.safetensors |
10.42GB (3 split files) |
Quantized and Converted Variants
→ Scroll horizontally to see all columns
| Added | Publisher | Format | Repository | Smallest VRAM tier (build, est. memory) |
|---|---|---|---|---|
| 2026-09-26 | unsloth | GGUF | unsloth/MiniMax-H3-GGUF | Q2_K 7.5GB (fits in 8GB VRAM) |
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
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.
Articles (the family’s own models first, then newest)
| Published | Model | Type | Article |
|---|---|---|---|
| 2026-09-26 | MiniMaxAI/MiniMax-H3 | Image, Video and Audio | MiniMax-H3 Audio-Visual Video Generation Model: 32GB+ VRAM, File List |
Repositories
Last updated 2026-09-27 (JST). The explanation at the top of this page was written with the help of AI from the primary sources it cites. The tables and lists under “Our Coverage and Data" are assembled by code from our article log.