MiniMax H3 Requirements: VRAM, GPU & File Sizes (2026)

MiniMax H3 is an open-weight model that generates video and audio simultaneously in a single pass, with weights released on August 3, 2026. Short answer: the minimum configuration recommended by ComfyUI (pruned INT8 + NVFP4 quantization) totals about 42.5GB of files on disk, with VRAM needs reported by the community at roughly 24GB (some reports suggest even 12GB can work with layer-wise offloading). A full bf16 setup, by contrast, totals about 123.6GB, and MiniMax's official SGLang example assumes a 4-GPU configuration. Below is a quick-reference table by quantization pattern. See also our requirements guides for GLM-5.2 and DeepSeek V4.

This table organizes published file sizes plus community and ComfyUI-team reports — it is not an official 'requirements table' from MiniMax (more on that below). MiniMax H3 was announced on July 31, 2026 by the Chinese AI company MiniMax, with weights released on August 3. It is hosted on Hugging Face as MiniMaxAI/MiniMax-H3, and the official ComfyUI distribution is Comfy-Org/MiniMax-H3. The architecture is a 33.1B-parameter dense single-stream omni transformer (13B of which is an AdaLN branch), paired with Qwen3-VL-32B as the text encoder.

Two H3-Base checkpoints were released: FL2VA (text-to-video / first-last-frame specification) and Ref2VA (image, video, and audio reference input). The preprocessing module H3-Context-IR and the 2K upscaling module H3-Regenerate-2K remain unreleased and are only available via MiniMax's hosted API. Output runs up to about 15 seconds at 24fps, 768p by default (2K via Regenerate-2K), with 32kHz stereo audio generated in the same pass. It supports 11 languages including Japanese (ar/zh/en/fr/de/it/ja/ko/pt/ru/es) and aspect ratios of 21:9/16:9/4:3/1:1/3:4/9:16. Listed inference frameworks are Diffusers, SGLang, vLLM, and ComfyUI.

The following are measured file sizes from the Hugging Face API for Comfy-Org/MiniMax-H3. FL2VA and Ref2VA share identical diffusion file sizes, split across quantization variants.

Pruned INT8 is about 40% smaller than standard INT8_convrot because the AdaLN curve tables are pre-computed. If disk space is a constraint, start with the pruned variant. For a broader look at running self-hosted open-weight models like this, see our Kimi K3 open-weights self-hosting guide.

- bf16: Maximum fidelity. Totals about 123.6GB on disk and points toward an 80GB-class GPU or multi-GPU setup. Best for quality checks and research.
- int8_convrot (non-pruned): A balance of quality and size, totaling about 67GB — a reasonable fit for a single 48GB-class GPU.
- pruned_int8_convrot: About 40% smaller thanks to pre-computed AdaLN curve tables. This is what ComfyUI's official workflow uses by default, bringing the minimum total to about 42.5GB.
- nvfp4_awq (text encoder only): A further-compressed format for Qwen3-VL-32B at 15.69GB, typically paired with pruned INT8 diffusion in the standard minimum configuration.

ComfyUI's official workflow (shared across T2V/I2V and R2V) recommends four files: minimax_h3_{fl2va|ref2va}_pruned_int8_convrot.safetensors, qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors, minimax_h3_video_vae_fp16.safetensors, and minimax_h3_audio_vae_fp32.safetensors. Adding 20.97GB + 15.69GB + 5.21GB + 0.61GB comes to about 42.5GB total. Note that using both FL2VA and Ref2VA adds roughly another 21GB each for the extra diffusion file.

MiniMax has not published an explicit VRAM requirements table. The following is a rough guide based on community testing and comments from the ComfyUI team — treat it as unofficial reference, not a spec.

- 24GB-class (e.g. RTX 4090): Reportedly workable with pruned INT8 + NVFP4 plus layer-wise offloading — a realistic baseline.
- 12GB-class: The ComfyUI team has commented that 12GB may work given ample system RAM and fast storage, though speed and stability are not guaranteed.
- 48GB-class (e.g. RTX 6000 Ada): A reasonable target for the non-pruned int8_convrot configuration with lighter offload demands.
- 80GB-class / multi-GPU (e.g. H100): The target for full bf16. The official SGLang example is built around a 4-GPU setup.

There is little official or community verification specific to MiniMax H3 on Apple Silicon at this time. In theory, a Mac with a large unified memory pool (say, 128GB or more, as on some Mac Studio configurations) could hold a lower-precision quantization variant in memory, but real-world confirmation of ComfyUI's MPS backend performance with this model is scarce. It's best to hold off on planning a deployment around Apple Silicon until concrete test reports appear.

MiniMax H3 is released under the MiniMax H3 Community License Agreement. Non-commercial use is unrestricted. Commercial use is permitted, with attribution, for organizations whose trailing 12-month revenue is under US$20 million. Organizations above that threshold may need a separate license agreement with MiniMax, so check the license terms carefully before deployment.

- Can: Generate up to roughly 15 seconds of video at 24fps and 768p by default, with synchronized 32kHz stereo audio; accept text or first/last-frame specification (FL2VA) or image/video/audio references (Ref2VA); support 11 languages; output aspect ratios from 21:9 to 9:16.
- Cannot (with the released weights): 2K upscaling (H3-Regenerate-2K is unreleased and hosted-API only) or advanced input conditioning via H3-Context-IR (also unreleased). Both require MiniMax's hosted service.

- Out of VRAM during load: Switch to the minimum pruned_int8_convrot + nvfp4_awq configuration first, then enable layer-wise offloading if it's still not enough.
- Unstable even with offloading: System RAM and storage speed may be the bottleneck — the ComfyUI team's 12GB VRAM scenario assumes ample system RAM and fast storage.
- Generation is slow or memory-constrained: Try lowering resolution below 768p or shortening clip length, and switch the text encoder to a lighter quantization such as nvfp4_awq.
- Disk fills up when using both FL2VA and Ref2VA: Each adds about 21GB of diffusion files under pruned INT8, so consider installing only the checkpoint you actually need.