Which GPU do you need for MiniMax H3?

Short answer: MiniMax H3 is large. The transformer is 33B parameters, and the text encoder is the full Qwen3-VL-32B model. In BF16 the transformer alone is about 66 GB on Hugging Face and the encoder about 51.5 GB, so no single consumer card holds it whole. MiniMax’s model card gives no VRAM figure. The diffusers documentation gives three tiers:

  • 24–32 GB cards (RTX 4090, RTX 5090): quantize to int8 and stream the transformer’s blocks from CPU RAM.
  • One 80 GB card (A100 80G, A800 80G): load the components on and off the GPU as needed.
  • Two 80 GB cards: run in full BF16.

On the smaller cards, host RAM matters as much as VRAM: diffusers says to “expect the weights to live in host RAM: around 75 GB of it at int8” (diffusers MiniMax H3 docs, read 2026-11-04). Read the licence section below before you use H3.

What is open, and what is not

The open release is the H3-Base model at 768p. MiniMax publishes it as two task-specific checkpoints (MiniMax H3 model card, read 2026-11-04):

Component Status
H3-Base FL2VA (text to video+audio, first/last frame) Open, BF16, CFG-distilled
H3-Base Ref2VA (reference images, video and audio) Open, BF16, CFG-distilled
H3-Regenerate-2K (768p to 2K) Not open: “this module is not yet open-sourced”
H3-Context-IR (prompt preprocessing) API only: “it is not included in this open-source release”
Sparse attention “will be released in a future update”

The model card says local output defaults to “the shorter side is set to 768 pixels”. Clips are 4–15 seconds at 24 FPS, with native 32 kHz stereo audio generated together with the video. The ComfyUI docs say “2K is a hosted-service output and needs a separate upscale pass in ComfyUI” (ComfyUI MiniMax H3 tutorial, read 2026-11-04).

Which card for which setup

The setups below come from the diffusers, SGLang and vLLM documentation (read 2026-11-04). The Deploy page lists the cards you can start now, with each card’s vCPU and RAM per GPU.

VRAM class Cards on CloudGPU What the docs describe
24–32 GB RTX 4090, RTX 4090D, RTX 5090, RTX 5090D diffusers: “quantize the two large components to int8 as they load and stream the transformer’s blocks from CPU RAM”. SGLang on one RTX 4090: “stream the DiT and text encoder and quantize DiT linear layers online with convrot_int8”; “GPU peak stayed about 18 GB”
48 GB RTX 4090 48G, RTX 4090D 48G, RTX A6000 More headroom for the int8 transformer and encoder. diffusers: “Two 48 GB cards do the same with the int8 loading” as two 80 GB cards in BF16
80–96 GB A100 80G, A800 80G, RTX 6000D 83G, RTX PRO 6000 96G, RTX PRO 6000 Max-Q 96G diffusers: “On one 80 GB card, register the components in a ComponentsManager and let it move them on and off the accelerator”
2 × 80 GB Two A100 80G or A800 80G diffusers: “Two 80 GB cards run full bfloat16 this way”

SGLang’s notes show that 80 GB is not enough to keep everything resident: “Pure Ulysses4 could not keep the full pipeline resident on 80 GB H100s” (SGLang MiniMax H3 cookbook, read 2026-11-04).

Host RAM is part of the sizing

Offloading moves the weights into system memory, so check the RAM per GPU shown on /deploy before you choose a card. The published numbers:

  • diffusers, 12–16 GB setup: “around 75 GB of it at int8”.
  • SGLang: a 2 × RTX 5090 recipe “was validated on 2× RTX 5090 (32 GB each) and a 377 GiB host; use a 384 GiB-class machine”. A single RTX 5090 desktop run used 32 GB VRAM and 60 GB host RAM at 864×480.
  • vLLM: “Use a 384 GiB-class host for the pinned model shards and offload buffers” (vLLM MiniMax H3 recipe, read 2026-11-04).

If a run is killed with no CUDA error, you most likely ran out of host RAM, not VRAM.

Which files to download

The Comfy-Org repack offers smaller files than the official BF16 shards. Sizes are from the Hugging Face API, read 2026-11-04.

File Size Notes
minimax_h3_fl2va_bf16 (also ref2va_) 66.28 GB full transformer
minimax_h3_fl2va_int8_convrot 34.04 GB int8
minimax_h3_fl2va_pruned_bf16 40.23 GB AdaLN branches removed
minimax_h3_fl2va_pruned_int8_convrot 20.97 GB pruned, int8
minimax_h3_fl2va_pruned_fp8_scaled 20.96 GB pruned, FP8
minimax_h3_fl2va_pruned_w6a8 15.98 GB pruned, w6a8
qwen3vl_32b_minimax_h3_bf16 (text encoder) 51.51 GB
qwen3vl_32b_minimax_h3_int8_convrot 27.14 GB
qwen3vl_32b_minimax_h3_nvfp4_awq 15.69 GB “does not require Blackwell GPU to use”
minimax_h3_video_vae_fp16 / _int8_convrot 5.21 / 2.81 GB
minimax_h3_audio_vae_fp32 0.61 GB

The Comfy README says “For diffusion models prefer int8_convrot if you are able to use pytorch with cu130” and “fp8_scaled should only be used if you cannot use int8_convrot.” The pruned files can be smaller because the model card says the AdaLN branches (about 13B parameters) “do not need to be loaded for inference-only deployment”.

Published speed figures

MiniMax’s own card has no timing figures. SGLang’s cookbook (read 2026-11-04) reports these runs:

Card Setup Time
1 × RTX 4090 D 24 GB 1344×768, 107 frames, 20 steps, BF16 + FlashAttention 405.6 s
1 × RTX 4090 D 24 GB same, convrot_int8 + FA 303.3 s
1 × RTX 4090 D 24 GB same, convrot_int8 + Sage→Sol hybrid 163.8 s
1 × RTX 5090 desktop 864×480, 124 frames, 20 steps, sglang 112.2 s
1 × RTX 5090 desktop same, ComfyUI --fast-disk --cache-none 145.9 s

diffusers adds that “960x544 runs about 2.3x faster per step than the trained 1344x768”. To get timings for your own prompts, start with a short clip at 960x544, then scale up.

License

MiniMax H3 is released under the MiniMax H3 Community License Agreement, not an open-source licence. Read the LICENSE file before you use the weights or their outputs. Renting a GPU does not change the licence terms.

Cost

At 2026-11-05 07:55 (UTC+8): RTX 5090 $0.69/hour, RTX 4090 48G $0.69/hour, A100 80G $1.59/hour. Billing is per minute, rounded up, with a 1-minute minimum, and it stops when you destroy the instance. Weight downloads are billed time too, so download once and render in batches. Current rates are on /pricing.

Related: Which GPU for Wan video generation · AI short film production: which GPU for each step · Cloud GPU for AI film and video production

Frequently asked questions

How much VRAM does MiniMax H3 need?

MiniMax’s model card gives no VRAM figure. The diffusers docs describe 24–32 GB cards with int8 and block streaming from CPU RAM, one 80 GB card with components moved on and off the GPU, and two 80 GB cards for full bfloat16. SGLang measured a GPU peak of about 18 GB on one RTX 4090 with streaming and int8 (read 2026-11-04).

Can MiniMax H3 run on an RTX 5090?

Yes, with quantization and offload. SGLang reports a single RTX 5090 desktop run at 864×480, 124 frames and 20 steps that took 112.2 s with sglang and 145.9 s with ComfyUI, using 32 GB of VRAM and 60 GB of host RAM (read 2026-11-04).

Does the open MiniMax H3 release generate 2K video?

No. The open checkpoints are H3-Base at a 768-pixel short side. MiniMax says H3-Regenerate-2K “is not yet open-sourced”, and the ComfyUI docs say 2K needs a separate upscale pass (read 2026-11-04).

Why does host RAM matter for MiniMax H3?

On cards under 80 GB, the weights are streamed from system memory. The diffusers docs say to expect around 75 GB of host RAM at int8, and the SGLang and vLLM recipes for two consumer cards ask for a 384 GiB-class host. Check the RAM per GPU on the CloudGPU Deploy page before choosing a card.

Is MiniMax H3 open source?

MiniMax H3 is released under the MiniMax H3 Community License Agreement, not an open-source licence. Read the LICENSE file on the MiniMax H3 Hugging Face page before you use the weights or their outputs.

Sources (read 2026-11-04, Asia/Shanghai)

The Deploy page lists the cards you can start. Prices are on https://cloudgpu.app/pricing.

Last updated: 2026-11-04 (Asia/Shanghai).