π£ Patched-Metal / Specialized Runtime: Quantized DiT + Qwen3-VL Text Encoder β runs in
stable-diffusion.cpp(sd-server/sd-cli, Metal 4) or ComfyUI.
β Support Local AI Development: Help me acquire a Mac mini 24GB/32GB to optimize and test local models up to 80B parameters. Donate via Ko-fi β
Kenzan-MiniMax-H3-Image-20B-PinGrid β "20B DiT + 25B Qwen3-VL Text Encoder + 1.1GB VAE, Running Locally on 16 GB RAM." 37.47 GiB BF16 DiT compressed 4.46Γ to 8.40 GiB,
int8_convrot50-layer Qwen3-VL-32B unpacked and Pin-Grid quantized to 10.46 GiB, and native Apple Metal (stable-diffusion.cpp) delivering ~7.2s/step at 1024Γ1024 with zero OOMs.
Overview
This repository provides a complete, 16 GB Unified Memory-ready GGUF release of MiniMaxAI/MiniMax-H3-Image (from Comfy-Org/MiniMax-H3-Image) tuned for Apple Silicon (Mac mini M6 16 GB) and compatible with both stable-diffusion.cpp (sd-cli / sd-server with Metal) and ComfyUI + ComfyUI-GGUF.
Included Files
| Component | File | Size | Compression / Precision |
|---|---|---|---|
| Diffusion Transformer (20.11B, 50 blocks) | Kenzan-MiniMax-H3-Image-20B-PinGrid-Q3_K-M6-16G.gguf |
8.40 GiB | Pin-Grid Q3_K / Q4_0 / Q5_K / Q2_K (F32 norms & adaln) β 4.46Γ vs BF16 (37.47 GiB) |
Text/Vision Encoder (Qwen3-VL-32B L50) |
Kenzan-Qwen3VL-32B-L50-MiniMaxH3-PinGrid-Q2_K-M6-16G.gguf |
10.46 GiB | Dequantized from int8_convrot (25.28 GiB), Pin-Grid Q2_K/Q3_K/Q4_0 + Q8_0 visual.* for Image Edit |
Image VAE (MiniMaxH3ImageVAE) |
minimax_h3_image_vae_fp16.safetensors |
1.09 GiB | Original lossless FP16 |
| Quantization Recipes | recipes/convert_dit_to_gguf.py, recipes/convert_te_to_gguf.py |
β | Exact streaming GGUF conversion scripts |
Pin-Grid Architecture
- 20B DiT (
Kenzan-MiniMax-H3-Image-20B-PinGrid-Q3_K-M6-16G.gguf):token_refiner.*,final_layer.*:Q5_K- Entry & tail blocks
blocks.(0,1,48,49)+blocks.(40..47).attn.*+ peakmlp.fc2(8, 16, 24, 32, 40):Q4_0 - Mid-late
blocks.(20..47):Q3_K - Early MLP
blocks.(2..19):Q2_K adaln_t_table,adaln_proj,patch_proj, norms:F32
- 50-Layer
Qwen3-VL-32BText Encoder (Kenzan-Qwen3VL-32B-L50-MiniMaxH3-PinGrid-Q2_K-M6-16G.gguf):visual.*(for Image Edit):Q8_0model.embed_tokens:Q5_K- Entry & exit layers
0, 48, 49+k_proj/v_proj+ peakmlp.down_proj(8, 16, 24, 32, 40):Q4_0 - Mid-late
layers.(20..47):Q3_K, early MLPlayers.(1..19):Q2_K
Verified Benchmarks (Apple Mac mini M6, 16 GB Unified Memory, 1024Γ1024, 6 steps euler/simple)
| Backend | DiT Step Speed | VAE Decode | Peak RSS | Notes |
|---|---|---|---|---|
ComfyUI + ComfyUI-GGUF (PyTorch MPS) |
13.27 s/step |
5.36 s |
11.88 GiB |
Per-layer runtime dequant to FP16 |
stable-diffusion.cpp (sd-cli / sd-server, Metal 4) |
7.20 β 8.24 s/step |
2.43 β 2.59 s |
Fits 16 GB (--offload-to-cpu) |
1.8Γ faster DiT, 2.1Γ faster VAE |
Quick Start (stable-diffusion.cpp on Apple Silicon 16 GB)
CLI (sd-cli)
./sd-cli --diffusion-model Kenzan-MiniMax-H3-Image-20B-PinGrid-Q3_K-M6-16G.gguf --llm Kenzan-Qwen3VL-32B-L50-MiniMaxH3-PinGrid-Q2_K-M6-16G.gguf --vae minimax_h3_image_vae_fp16.safetensors --offload-to-cpu --flash-attn on -W 1024 -H 1024 --steps 6 --cfg-scale 1.0 --sampling-method euler --scheduler simple -p "Your prompt here" -o output.png
Persistent Server (sd-server, eliminates cold-start weight loading)
./sd-server --listen-ip 0.0.0.0 --listen-port 8189 --diffusion-model Kenzan-MiniMax-H3-Image-20B-PinGrid-Q3_K-M6-16G.gguf --llm Kenzan-Qwen3VL-32B-L50-MiniMaxH3-PinGrid-Q2_K-M6-16G.gguf --vae minimax_h3_image_vae_fp16.safetensors --offload-to-cpu --flash-attn on
Acknowledgements & Inspiration
- Inspired by Bonsai-2 & Prism ML: Deep gratitude and inspiration go to the creators of Bonsai-2 at Prism ML (huggingface.co/prism-ml) for pioneering extreme low-bit quantization and proving that frontier-class models can thrive on consumer unified memory.
- Special Thanks to Misha: Sincere gratitude to Misha for unwavering support and financial assistance, which made the hardware testing, calibration runs, and open-weights releases in the Kenzan series possible.
- Upstream Projects:
MiniMaxAI/MiniMax-H3-Image,Comfy-Org/MiniMax-H3-Image,leejet/stable-diffusion.cpp, andcity96/ComfyUI-GGUF.
Support the Project & Funding Goals
All Kenzan weights and quantization recipes are released openly and free of charge for the local AI community.
What Your Donations Directly Fund
- Upgrading to a Mac mini 24 GB / 32 GB (Apple Silicon): Breaking past the 16 GB ceiling to quantize and benchmark 70Bβ80B models.
- Cloud GPU Compute for Calibration Sweeps: Covering VRAM hours for
imatrixgeneration and layer sensitivity sweeps. - Open-Source Metal Kernels & Tooling: Funding continued development of
llama.cpp/stable-diffusion.cppMetal optimizations.
- Downloads last month
- 261
2-bit
Model tree for kabcnd/Kenzan-MiniMax-H3-Image-20B-PinGrid-GGUF
Base model
MiniMaxAI/MiniMax-H3