Files read on Hugging Face: 2026-10-24

SDXL Base 1.0 on Mac: files, RAM and published times

SDXL Base 1.0 text-to-image model with a UNet denoiser and two CLIP text encoders. The refiner is an optional second stage; the base model works standalone.

Stability AI · Updated 2026-10-24
Answer

SDXL Base 1.0 runs on a Mac with 256 GB or more. A published run on M3 Ultra · 256 GB used the Rapid-MLX 8-bit UNet build (source, 2026-09-05).

Smallest Mac with a published run
256 GBM3 Ultra · 256 GB · Rapid-MLX 8-bit UNet
Smallest complete download
2.80 GBQ4_K GGUF
Files read
2026-10-24Hugging Face

CreativeML Open RAIL++-M License. Commercial use: allowed. Excluded territories: none listed. Broad use grant with the use-based restrictions in Attachment A.

See the exact GGUF files and sizes

This is SDXL Base 1.0. The SDXL Refiner 1.0 is an optional second-stage model and is not included in these sizes.

Table 1

What do you download?

Only builds with every required component and an exact probed file size are shown.

Rapid-MLX 8-bit UNet

safetensors · 8-bit UNet (quantized at load) · 6.94 GB complete · Recommended

Mac routes: Rapid-MLX

Exact component files for Rapid-MLX 8-bit UNet
FileComponentRequiredDownloadSource
unet/diffusion_pytorch_model.fp16.safetensorsstabilityai/stable-diffusion-xl-base-1.0SDXL UNetditYes5.14 GBHugging FaceRead 2026-10-24
text_encoder/model.fp16.safetensorsstabilityai/stable-diffusion-xl-base-1.0CLIP ViT-L text encoder (inside single-file checkpoints)text_encoderNo0.25 GBHugging FaceRead 2026-10-24
text_encoder_2/model.fp16.safetensorsstabilityai/stable-diffusion-xl-base-1.0OpenCLIP ViT-bigG text encoder (inside single-file checkpoints)text_encoderNo1.39 GBHugging FaceRead 2026-10-24
vae/diffusion_pytorch_model.fp16.safetensorsstabilityai/stable-diffusion-xl-base-1.0VAE (inside single-file checkpoints)vaeNo0.17 GBHugging FaceRead 2026-10-24

Build sources: Rapid-MLX SDXL Base dogfood.

Q4_K GGUF

gguf · Q4_K · 2.80 GB complete

Mac routes: stable-diffusion.cpp

Exact component files for Q4_K GGUF
FileComponentRequiredDownloadSource
sd_xl_base_1.0_0_Q4_K.ggufkostakoff/stable-diffusion-xl-base-1.0-GGUFSDXL UNetditYes2.80 GBHugging FaceRead 2026-10-24

Build sources: SDXL GGUF card.

Q8_0 GGUF

gguf · Q8_0 · 4.18 GB complete

Mac routes: stable-diffusion.cpp

Exact component files for Q8_0 GGUF
FileComponentRequiredDownloadSource
sd_xl_base_1.0_0_Q8_0.ggufkostakoff/stable-diffusion-xl-base-1.0-GGUFSDXL UNetditYes4.18 GBHugging FaceRead 2026-10-24

Build sources: SDXL GGUF card.

Original single-file checkpoint

safetensors · Original weights · 6.94 GB complete

Mac routes: ComfyUI

Exact component files for Original single-file checkpoint
FileComponentRequiredDownloadSource
sd_xl_base_1.0.safetensorsstabilityai/stable-diffusion-xl-base-1.0SDXL UNetditYes6.94 GBHugging FaceRead 2026-10-24

Build sources: ComfyUI README.

Table 2

Does it fit your Mac's RAM?

Published names the exact measured machine. est. is a memory calculation. unknown means the run verdict is not established, not that the model fails.

Run verdicts by Mac memory tier
Mac RAMComplete buildDownloadRun verdictBasis
8 GBRapid-MLX 8-bit UNet8-bit UNet (quantized at load) · Rapid-MLX6.94 GBunknownunknown
16 GBRapid-MLX 8-bit UNet8-bit UNet (quantized at load) · Rapid-MLX6.94 GBunknownunknown
24 GBRapid-MLX 8-bit UNet8-bit UNet (quantized at load) · Rapid-MLX6.94 GBunknownunknown
32 GBRapid-MLX 8-bit UNet8-bit UNet (quantized at load) · Rapid-MLX6.94 GBunknownunknown
48 GBRapid-MLX 8-bit UNet8-bit UNet (quantized at load) · Rapid-MLX6.94 GBunknownunknown
64 GBRapid-MLX 8-bit UNet8-bit UNet (quantized at load) · Rapid-MLX6.94 GBunknownunknown
96 GBRapid-MLX 8-bit UNet8-bit UNet (quantized at load) · Rapid-MLX6.94 GBunknownunknown
128 GBRapid-MLX 8-bit UNet8-bit UNet (quantized at load) · Rapid-MLX6.94 GBunknownunknown

How does it rank in blind votes?

No independent score yet.

Table 3

What published Mac times are available?

Third-party results keep their original machine and settings. ModelFit does not run these benchmarks.

Published Mac generation measurements
Machine and RAMSettingsAppResultSource
Apple M3 Ultra · 256 GB RAM1024x1024 · 30 steps · 8-bit UNet (quantized at load) · Guidance 5.0, seed 42, tiled VAE, macOS 26.5.2, MLX 0.32.2. Time is cold model construction plus generation. Peak is MLX peak allocation.Rapid-MLXSuccess61.7 s12.42 GiBRapid-MLX SDXL Base dogfoodPublished 2026-09-05 · Grade A

Which Mac software routes are listed?

  • ComfyUI

    Official

    Formats: safetensors, gguf. Memory policy: component stages stay resident.

    On Apple Silicon ComfyUI runs on PyTorch MPS, not MLX. GGUF files need the separate city96/ComfyUI-GGUF custom node. Resident size is the sum of the weights, an upper bound: ComfyUI can unload models between stages and publishes no per-model Mac peak.

    Sources: ComfyUI README · ComfyUI system requirements · ComfyUI CLI arguments · ComfyUI-GGUF README

  • Diffusers

    Official

    Formats: safetensors, gguf. Memory policy: component stages stay resident.

    Runs on the PyTorch mps device on Mac and CUDA on NVIDIA. GGUF loads through from_single_file on the model class, not the pipeline. Resident size is the sum of the weights, an upper bound.

    Sources: Diffusers MPS guide · Diffusers GGUF guide

  • Draw Things

    Official

    Formats: safetensors. Memory policy: component stages stay resident.

    Native offline app for Mac. It picks a weight-streaming tier from physical memory. No official Mac timing or peak-memory table was found, so there is no run verdict here.

    Sources: Draw Things on the App Store · Draw Things App Store listing data · Draw Things model importer

  • Rapid-MLX

    Community

    Formats: safetensors, mlx. Memory policy: component stages stay resident.

    MLX-native image server with an OpenAI-compatible images endpoint. SDXL Base quantizes the UNet to 8-bit at load; SD 3.5 Large uses a 4-bit MLX checkpoint.

    Sources: Rapid-MLX SDXL Base dogfood · Rapid-MLX SD 3.5 Large dogfood

  • stable-diffusion.cpp

    Official

    Formats: gguf, safetensors. Memory policy: component stages stay resident.

    C/C++ runner with Metal on Mac and CUDA on NVIDIA. It reads GGUF and safetensors, and its docs keep a page per model family. Resident size is the sum of the weights, an upper bound.

    Sources: stable-diffusion.cpp README, backends

What does the official configuration say?

"No numeric RAM or VRAM minimum is given. The card suggests CPU offloading when limited by GPU VRAM."

Source: SDXL Base model card · Read 2026-10-24.

What known issues should you check?

  • Known issue 1

    ComfyUI documents that running the VAE in fp16 might cause black images. Individual SDXL black-image reports on Macs exist but are not general rules.

    Source: ComfyUI CLI arguments · Read 2026-10-24

Frequently asked questions

How fast is SDXL on a Mac?

Rapid-MLX with an 8-bit UNet made a 1024x1024 image in 30 steps in 61.72 s, including cold model construction, on an M3 Ultra with 256 GiB of memory. The MLX peak allocation was 12.42 GiB.

Do I need the SDXL refiner?

No. The card says the base model can be used as a standalone module. The SDXL Refiner 1.0 is an optional second-stage model and is not included in the sizes on this page.

Why do some SDXL images come out black on a Mac?

ComfyUI documents that running the VAE in fp16 might cause black images. Individual Mac reports exist, but they are single issue reports, not general rules, so a black image alone does not mean the Mac ran out of memory.

Sources