MiniMax H3-Image Runs on Consumer GPUs as Small as 8GB
GGUF quantizations of MiniMax-H3-Image now run on 8 to 24 GB consumer GPUs Five quants from Q2_K (6.88 GB) to Q5_K_M (14.1 GB), with Q4_K_M as the recommended sweet spot Sensitive layers (patchifier, PDD head, AdaLN basis) kept in FP32 to avoid grid artifacts Loads natively in stock ComfyUI-GGUF
Takeaways
- GGUF quantizations of MiniMax-H3-Image now run on 8 to 24 GB consumer GPUs
- Five quants from Q2_K (6.88 GB) to Q5_K_M (14.1 GB), with Q4_K_M as the recommended sweet spot
- Sensitive layers (patchifier, PDD head, AdaLN basis) kept in FP32 to avoid grid artifacts
- Loads natively in stock ComfyUI-GGUF with embedded 6-step PDD schedule metadata
- Base MiniMax H3-Image is a 33.1B DiT producing 4K images in 6 steps, no CFG
- Requires separate 24-channel Image VAE and Qwen3-VL-32B text encoder; license excludes US, EU, UK
Community publisher Abiray has released five GGUF files for MiniMax-H3-Image, built from Comfy-Org’s pruned checkpoint. The collection ranges from a 6.88 GB 2-bit build, designed to fit the image transformer within an 8 to 10 GB GPU tier, to a 14.1 GB 5-bit version that the publisher describes as visually comparable to its BF16 source checkpoint.
GGUF packages model tensors at reduced precision, cutting storage and GPU memory at some cost to output quality. These releases make H3-Image’s diffusion transformer practical on consumer hardware and load through stock ComfyUI-GGUF without a custom patch. The complete pipeline still requires a large text encoder and a dedicated VAE, so the file size alone does not represent peak memory use.
Two cuts bring 33B down
MiniMax introduced the H3-Image release as an open-weight model for text-to-image generation and reference-based editing. Its 33.1 billion-parameter diffusion transformer derives from the company’s H3 video architecture, with the audio branch removed and the decoder redesigned for still images.
The dense transformer has 50 layers, a hidden size of 5,376, 56 attention heads and a feed-forward width of 14,336. It uses 3D multimodal rotary position embeddings to track spatial and multimodal relationships. The model can produce images up to 4096 × 4096, accept as many as 16 reference images and complete its denoising process in six sampling steps.
Those six passes come from distillation, which trains the model around an abbreviated noise-removal schedule. H3-Image also omits classifier-free guidance, a common second inference path used to strengthen prompt adherence. MiniMax reports that Standard mode runs 1.5 to 2 times as fast as its comparison set, while Highspeed mode reaches 3 to 4 times the speed. Those figures are vendor measurements and will vary with hardware, resolution and workflow.
Comfy-Org’s structurally pruned BF16 checkpoint reduces the effective parameter count to about 20 billion before quantization. Structural pruning removes model capacity, while quantization stores the remaining weights with fewer bits. Combining both techniques produces the following files:
Reporting is indexed from AlphaSignal. Rights remain with the original publisher and cited sources.
Read original report
