Files read on Hugging Face: 2026-10-24

Qwen-Image-2.1 on Mac: files, RAM and published times

Base Qwen-Image-2.1 text-to-image and editing model with a 7B visual generation component and a Qwen3-VL text encoder. The Qwen Research License allows non-commercial use only.

Qwen · Updated 2026-10-24
Answer

Qwen-Image-2.1 runs on a Mac with 48 GB or more. A published run on M5 Pro · 48 GB used the BF16 safetensors build (source, 2026-10-11).

Smallest Mac with a published run
48 GBM5 Pro · 48 GB · BF16 safetensors
Smallest complete download
10.70 GBQ4_K_M GGUF
Files read
2026-10-24Hugging Face

Qwen Research License Agreement. Commercial use: not allowed. Excluded territories: none listed. Commercial use needs a separate commercial license from Qwen.

See the exact GGUF files and sizes

This is base Qwen-Image-2.1, not Qwen-Image-2.1-Turbo.

Table 1

What do you download?

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

BF16 safetensors

safetensors · BF16 · 33.12 GB complete · Recommended

Mac routes: Diffusers, Diffusers with staged loading

Exact component files for BF16 safetensors
FileComponentRequiredDownloadSource
transformer/diffusion_pytorch_model-00001-of-00002.safetensorsQwen/Qwen-Image-2.1Qwen-Image-2.1 transformerditYes9.97 GBHugging FaceRead 2026-10-24
transformer/diffusion_pytorch_model-00002-of-00002.safetensorsQwen/Qwen-Image-2.1Qwen-Image-2.1 transformerditYes4.26 GBHugging FaceRead 2026-10-24
text_encoder/model-00001-of-00004.safetensorsQwen/Qwen-Image-2.1Qwen3-VL text encodertext_encoderYes5.00 GBHugging FaceRead 2026-10-24
text_encoder/model-00002-of-00004.safetensorsQwen/Qwen-Image-2.1Qwen3-VL text encodertext_encoderYes4.92 GBHugging FaceRead 2026-10-24
text_encoder/model-00003-of-00004.safetensorsQwen/Qwen-Image-2.1Qwen3-VL text encodertext_encoderYes4.92 GBHugging FaceRead 2026-10-24
text_encoder/model-00004-of-00004.safetensorsQwen/Qwen-Image-2.1Qwen3-VL text encodertext_encoderYes2.70 GBHugging FaceRead 2026-10-24
vae/diffusion_pytorch_model.safetensorsQwen/Qwen-Image-2.1VAEvaeYes1.35 GBHugging FaceRead 2026-10-24

Build sources: Qwen-Image-2.1 model card · How to run Qwen-Image-2.1 on a 48 GB Mac without swap.

Q4_K_M GGUF

gguf · Q4_K_M · 10.70 GB complete

Mac routes: stable-diffusion.cpp

Exact component files for Q4_K_M GGUF
FileComponentRequiredDownloadSource
qwen-image-2.1-Q4_K_M.ggufunsloth/Qwen-Image-2.1-GGUFQwen-Image-2.1 transformerditYes4.20 GBHugging FaceRead 2026-10-24
Qwen3-VL-8B-Instruct-UD-Q4_K_XL.ggufunsloth/Qwen3-VL-8B-Instruct-GGUFQwen3-VL text encodertext_encoderYes5.15 GBHugging FaceRead 2026-10-24
vae/diffusion_pytorch_model.safetensorsQwen/Qwen-Image-2.1VAEvaeYes1.35 GBHugging FaceRead 2026-10-24

Build sources: Qwen-Image-2.1 GGUF card · stable-diffusion.cpp Qwen Image 2.1 guide.

Q8_0 GGUF

gguf · Q8_0 · 14.14 GB complete

Mac routes: stable-diffusion.cpp

Exact component files for Q8_0 GGUF
FileComponentRequiredDownloadSource
qwen-image-2.1-Q8_0.ggufunsloth/Qwen-Image-2.1-GGUFQwen-Image-2.1 transformerditYes7.64 GBHugging FaceRead 2026-10-24
Qwen3-VL-8B-Instruct-UD-Q4_K_XL.ggufunsloth/Qwen3-VL-8B-Instruct-GGUFQwen3-VL text encodertext_encoderYes5.15 GBHugging FaceRead 2026-10-24
vae/diffusion_pytorch_model.safetensorsQwen/Qwen-Image-2.1VAEvaeYes1.35 GBHugging FaceRead 2026-10-24

Build sources: Qwen-Image-2.1 GGUF card · stable-diffusion.cpp Qwen Image 2.1 guide.

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 GBBF16 safetensorsBF16 · Diffusers with staged loading33.12 GBDoes not fitest.
16 GBQ4_K_M GGUFQ4_K_M · stable-diffusion.cpp10.70 GBunknownunknown
24 GBBF16 safetensorsBF16 · Diffusers with staged loading33.12 GBunknownunknown
32 GBBF16 safetensorsBF16 · Diffusers with staged loading33.12 GBunknownunknown
48 GBBF16 safetensorsBF16 · Diffusers with staged loading33.12 GBRuns · PublishedPublishedApple M5 Pro · 48 GB RAM
64 GBBF16 safetensorsBF16 · Diffusers with staged loading33.12 GBRuns · Published on M5 Pro · 48 GBPublishedApple M5 Pro · 48 GB RAM
96 GBBF16 safetensorsBF16 · Diffusers with staged loading33.12 GBRuns · Published on M5 Pro · 48 GBPublishedApple M5 Pro · 48 GB RAM
128 GBBF16 safetensorsBF16 · Diffusers with staged loading33.12 GBRuns · Published on M5 Pro · 48 GBPublishedApple M5 Pro · 48 GB RAM

How does it rank in blind votes?

#19 of 82 in text to image blind votes, from 11,441 votes.

Quality: blind human votes from qwen-image-2.1 (arena.ai), CC BY 4.0, snapshot 2026-10-07. Scores are for the Arena checkpoint, not your local quantized build.

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 M5 Pro · 48 GB RAM1024x1024 · 20 steps · BF16 · Seed 7, torch 2.14.0, diffusers 0.41.0.dev0, one model in memory per stage. Two runs finished in 94 s and 124 s; the slower is shown. The footprint peaked at 19.0 GB while encoding.Diffusers with staged loadingSuccess124 s19.0 GBHow to run Qwen-Image-2.1 on a 48 GB Mac without swapPublished 2026-10-11 · Grade A

Which Mac software routes are listed?

  • 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

  • Diffusers with staged loading

    Community

    Formats: safetensors. Memory policy: components can be unloaded between stages.

    Loads one large model at a time (text encoder, then transformer, then VAE) with the stageload library. In the same test, eager loading and model CPU offload both went into swap on a 48 GB Mac.

    Sources: How to run Qwen-Image-2.1 on a 48 GB Mac without swap

  • mflux

    Community · 0.23.0

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

    MLX-native runner. It reads the original Hugging Face weights and quantizes them while loading (--quantize 3, 4, 5, 6 or 8). --low-ram lowers memory use at a performance cost.

    Install with uv

    uv tool install --upgrade mflux

    Sources: mflux README, install · mflux pyproject · mflux common README, quantization

  • 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?

"The card text-to-image example uses 2048 x 2048 and 40 steps. It gives no minimum VRAM or Mac RAM figure."

Source: Qwen-Image-2.1 model card · Read 2026-10-24.

What known issues should you check?

Frequently asked questions

What does Qwen-Image-2.1 need on a Mac?

One published run used BF16 with staged loading on an M5 Pro with 48 GB. A 1024x1024 image in 20 steps took 94 s and 124 s, with the footprint peaking at 19.0 GB. Eager loading and model CPU offload went into swap on the same Mac.

Can I use Qwen-Image-2.1 commercially?

Not without a separate license. The Qwen Research License Agreement grants use for non-commercial purposes only, and commercial use needs a separate commercial license from Qwen. The terms apply to the base Qwen-Image-2.1 model listed here.

Does Qwen-Image-2.1 have a GGUF build?

Yes. Unsloth publishes Q8_0 and Q4_K_M files for the denoiser only, so they need the VAE and the Qwen3-VL text encoder alongside them. The card says to run them with Unsloth Desktop or stable-diffusion.cpp.

Sources