← All guides

MiniMax H3 Image VRAM (2026): The Encoder Is the Big File

ComfyUI added MiniMax H3 Image on 10 November 2026, and GGUF files followed the same day. The model card lists a 33.1B image transformer, but that is not the number that sets your VRAM. We read every file size from the Hugging Face API on 12 November 2026.

Bottom Line

  • The text encoder is the biggest file, not the image model. The smallest ComfyUI set is 29.4 GB: a 12.55 GB image model, a 15.69 GB text encoder and a 1.17 GB VAE.
  • 24GB is the comfortable tier. Every single file of the small set fits on the card with room for 4K output.
  • 16GB should work with offload, unproven. ComfyUI loads the encoder and the image model in turn. The GGUF card says Q4_K_M needs “16 GB VRAM”, but it counts the image model only. Nobody had published a VRAM reading by 12 November 2026.
  • Read the license first. It excludes the United States, the EU, the UK and South Korea. People there need a separate license from MiniMax.
Your cardMiniMax H3 Image?Files to use
12GBPossible, slowQ3_K_M GGUF (8.91 GB) + NVFP4 encoder with offload
16GBLikely, with offloadQ4_K_M GGUF (11.57 GB) or w4a8 (12.55 GB) + NVFP4 encoder (15.69 GB)
24GBYesQ5_K_M (14.08 GB) or int8 (20.98 GB) + NVFP4 encoder
32GB+Yesint8 image model + int8 encoder (49.29 GB set)

Plan on 32GB of system RAM at minimum for the 29.4 GB set; 64GB if you also load the optional prompt expander.

What MiniMax H3 Image Is

From the MiniMax model card and the Comfy blog (10 November 2026):

PartValue
Image model (DiT)33.1B parameters, dense, 50 layers, BF16 weights
Text encoderThe full Qwen3-VL text encoder of MiniMax H3, unchanged
Sampling6 steps, guidance 1, no negative prompt (fixed by the checkpoint)
OutputNative up to 4096×4096 (1:1), 5472×3072 (16:9), 6272×2688 (21:9)
Default in ComfyUI2048×2048 for text-to-image
EditingSame checkpoint; connect up to several reference images
Prompt lengthUp to 4,096 tokens

It is not the H3 video model. Both share one text encoder, so an existing H3 setup already has that file.

Every File Size

Read from the Hugging Face API on 12 November 2026. GB means 10^9 bytes.

Official MiniMax repo (BF16, unpruned): transformer 66.28 GB, text encoder 66.73 GB, VAE 2.34 GB. That is 135 GB before quantization.

Comfy-Org repo (pruned image model):

FileSize (GB)
Image model, pruned BF1640.23
Image model, pruned int820.98
Image model, pruned w4a812.55
Text encoder (Qwen3-VL 32B), int827.14
Text encoder (Qwen3-VL 32B), NVFP415.69
Prompt expander (Qwen3.5 9B), w4a87.34
VAE, fp161.17

Abiray GGUF repo (image model only, from the pruned BF16 file):

QuantSize (GB)Card’s VRAM advice
Q2_K6.888-10 GB
Q3_K_M8.9112 GB
Q4_K_M11.5716 GB
Q5_K_M14.0816-24 GB

The Arithmetic Nobody Prints

The GGUF card’s “16 GB VRAM” is for the image model alone. A working setup adds the encoder and the VAE:

SetImage modelEncoderVAETotal (GB)
Q3_K_M GGUF8.9115.691.1725.77
Q4_K_M GGUF11.5715.691.1728.43
Comfy w4a812.5515.691.1729.41
Comfy int820.9827.141.1749.29

At every matched precision, the encoder is larger than the 4-bit image model. On 16GB, the encoder sets the floor, the same pattern as the H3 video model.

One more unknown: pruning. Every ComfyUI and GGUF file comes from a pruned image model. The pruned BF16 file is 40.23 GB against 66.28 GB for MiniMax’s transformer. Nobody has published what the pruning costs in image quality.

How to Run It in ComfyUI

From the Comfy blog:

  1. Update ComfyUI to the latest version.
  2. Download three files: an image model, the Qwen3-VL text encoder and the VAE.
  3. Load the MiniMax H3 Image template from the Templates panel.
  4. Write a long, complete prompt. To edit, connect images to image_1 onward and connect the VAE.

On 16GB, start at the 2048×2048 default. The 4096×4096 output needs far more activation memory; test it after the default works.

The License Catch

The MiniMax-H3-Image Community License grants rights only in the “Applicable Territory”: worldwide, excluding the European Union, the United Kingdom, the Republic of Korea and the United States. Use outside that territory is “not authorized”. Inside it, a business with more than $1 million in annual revenue, or anyone offering a hosted service, needs prior written authorization from MiniMax.

That is why this page links no hardware to buy: much of this site’s audience cannot use the model under these terms.

Sources

  • MiniMaxAI/MiniMax-H3-Image model card, LICENSE and file list (read 2026-11-12).
  • Comfy blog, “MiniMax H3 Image is now available in ComfyUI”, Rob, 10 November 2026.
  • Hugging Face API: Comfy-Org/MiniMax-H3-Image (created 2026-11-08) and Abiray/MiniMax-H3-Image-GGUF (created 2026-11-10) file sizes and card, read 2026-11-12.

See Also

Sizing a card for image and video models? The tested hardware list groups GPUs by VRAM tier.

Read next

Best Qwen Model for 16GB VRAM (2026): Qwen3.8 27B Wins
Best Qwen model for 16GB VRAM: Qwen3.8-27B at 3.5 bits (10.96 GiB). The Q4_K_M 27B files do not fit. Byte-exact GGUF sizes, KV math, and the best Qwen for 8, 12, 24 and 32GB.
MiniMax H3 Local Requirements: Yes, 16GB VRAM Works (2026)
MiniMax H3 is a 33B audio-video model plus a Qwen3-VL-32B text encoder. A 16GB card runs ComfyUI's pruned INT8 files with offload. 24GB holds the video model. System RAM and the license territory matter more than most guides say.
Can I Run LTX-2.5 Locally? VRAM by GPU (2026)
LTX-2.5 is a 22B video model plus a 12B Gemma 4 text encoder. The official INT8 files need a 24GB card. 16GB works with NVFP4 or GGUF and weight streaming. The '16GB minimum' leans on cloud text encoding. Ollama does not run it.
Can I Run Nemotron 3.5 Transcribe Locally? Yes, on 16GB
Nemotron 3.5 Transcribe is NVIDIA's 34B (4.3B active) speech-to-text model. Byte-exact GGUF sizes, what fits 16GB, 24GB and 32GB cards and Macs, and the two runtimes that load it.