Files read on Hugging Face: 2026-10-24

FLUX.2 [dev] on Mac: files, RAM and published times

32 billion parameter rectified flow transformer for text-to-image, editing and combining images, with a 24B Mistral Small text encoder. Model use is non-commercial; outputs may be used commercially within the license limits.

Black Forest Labs · Updated 2026-10-24
Answer

FLUX.2 [dev] runs on a Mac with 96 GB or more. A published run on M2 Ultra · 96 GB used the BF16 safetensors build (source, 2026-02-11).

Smallest Mac with a published run
96 GBM2 Ultra · 96 GB · BF16 safetensors
Smallest complete download
34.75 GBQ4_K_M GGUF
Files read
2026-10-24Hugging Face

FLUX Non-Commercial License. Commercial use: restricted. Excluded territories: none listed. Model use is non-commercial. Output may be used commercially except as expressly prohibited, and may not train, fine-tune or distill a competing FLUX model.

See the exact GGUF files and sizes

Table 1

What do you download?

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

BF16 safetensors

safetensors · BF16 · 112.81 GB complete · Recommended

Mac routes: Diffusers, flux-2-swift-mlx

Exact component files for BF16 safetensors
FileComponentRequiredDownloadSource
transformer/diffusion_pytorch_model-00001-of-00007.safetensorsblack-forest-labs/FLUX.2-devFLUX.2 transformerditYes9.94 GBHugging FaceRead 2026-10-24
transformer/diffusion_pytorch_model-00002-of-00007.safetensorsblack-forest-labs/FLUX.2-devFLUX.2 transformerditYes9.89 GBHugging FaceRead 2026-10-24
transformer/diffusion_pytorch_model-00003-of-00007.safetensorsblack-forest-labs/FLUX.2-devFLUX.2 transformerditYes9.81 GBHugging FaceRead 2026-10-24
transformer/diffusion_pytorch_model-00004-of-00007.safetensorsblack-forest-labs/FLUX.2-devFLUX.2 transformerditYes9.81 GBHugging FaceRead 2026-10-24
transformer/diffusion_pytorch_model-00005-of-00007.safetensorsblack-forest-labs/FLUX.2-devFLUX.2 transformerditYes9.81 GBHugging FaceRead 2026-10-24
transformer/diffusion_pytorch_model-00006-of-00007.safetensorsblack-forest-labs/FLUX.2-devFLUX.2 transformerditYes9.81 GBHugging FaceRead 2026-10-24
transformer/diffusion_pytorch_model-00007-of-00007.safetensorsblack-forest-labs/FLUX.2-devFLUX.2 transformerditYes5.36 GBHugging FaceRead 2026-10-24
text_encoder/model-00001-of-00010.safetensorsblack-forest-labs/FLUX.2-devMistral Small 3.2 text encodertext_encoderYes4.88 GBHugging FaceRead 2026-10-24
text_encoder/model-00002-of-00010.safetensorsblack-forest-labs/FLUX.2-devMistral Small 3.2 text encodertext_encoderYes4.78 GBHugging FaceRead 2026-10-24
text_encoder/model-00003-of-00010.safetensorsblack-forest-labs/FLUX.2-devMistral Small 3.2 text encodertext_encoderYes4.89 GBHugging FaceRead 2026-10-24
text_encoder/model-00004-of-00010.safetensorsblack-forest-labs/FLUX.2-devMistral Small 3.2 text encodertext_encoderYes4.78 GBHugging FaceRead 2026-10-24
text_encoder/model-00005-of-00010.safetensorsblack-forest-labs/FLUX.2-devMistral Small 3.2 text encodertext_encoderYes4.78 GBHugging FaceRead 2026-10-24
text_encoder/model-00006-of-00010.safetensorsblack-forest-labs/FLUX.2-devMistral Small 3.2 text encodertext_encoderYes4.89 GBHugging FaceRead 2026-10-24
text_encoder/model-00007-of-00010.safetensorsblack-forest-labs/FLUX.2-devMistral Small 3.2 text encodertext_encoderYes4.78 GBHugging FaceRead 2026-10-24
text_encoder/model-00008-of-00010.safetensorsblack-forest-labs/FLUX.2-devMistral Small 3.2 text encodertext_encoderYes4.78 GBHugging FaceRead 2026-10-24
text_encoder/model-00009-of-00010.safetensorsblack-forest-labs/FLUX.2-devMistral Small 3.2 text encodertext_encoderYes4.89 GBHugging FaceRead 2026-10-24
text_encoder/model-00010-of-00010.safetensorsblack-forest-labs/FLUX.2-devMistral Small 3.2 text encodertext_encoderYes4.57 GBHugging FaceRead 2026-10-24
vae/diffusion_pytorch_model.safetensorsblack-forest-labs/FLUX.2-devFLUX.2 VAEvaeYes0.34 GBHugging FaceRead 2026-10-24

Build sources: flux-2-swift-mlx model registry · FLUX.2-dev model card.

Q4_K_M GGUF

gguf · Q4_K_M · 34.75 GB complete

Mac routes: ComfyUI, stable-diffusion.cpp

Exact component files for Q4_K_M GGUF
FileComponentRequiredDownloadSource
flux2-dev-Q4_K_M.ggufcity96/FLUX.2-dev-ggufFLUX.2 transformerditYes20.08 GBHugging FaceRead 2026-10-24
Mistral-Small-3.2-24B-Instruct-2506-Q4_K_M.ggufunsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUFMistral Small 3.2 text encodertext_encoderYes14.33 GBHugging FaceRead 2026-10-24
ae.safetensorsblack-forest-labs/FLUX.2-devFLUX.2 VAEvaeYes0.34 GBHugging FaceRead 2026-10-24

Build sources: FLUX.2-dev GGUF card · stable-diffusion.cpp FLUX.2 guide.

Q8_0 GGUF

gguf · Q8_0 · 49.67 GB complete

Mac routes: ComfyUI, stable-diffusion.cpp

Exact component files for Q8_0 GGUF
FileComponentRequiredDownloadSource
flux2-dev-Q8_0.ggufcity96/FLUX.2-dev-ggufFLUX.2 transformerditYes35.00 GBHugging FaceRead 2026-10-24
Mistral-Small-3.2-24B-Instruct-2506-Q4_K_M.ggufunsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUFMistral Small 3.2 text encodertext_encoderYes14.33 GBHugging FaceRead 2026-10-24
ae.safetensorsblack-forest-labs/FLUX.2-devFLUX.2 VAEvaeYes0.34 GBHugging FaceRead 2026-10-24

Build sources: FLUX.2-dev GGUF card · stable-diffusion.cpp FLUX.2 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 · flux-2-swift-mlx112.81 GBDoes not fitest.
16 GBBF16 safetensorsBF16 · flux-2-swift-mlx112.81 GBDoes not fitest.
24 GBBF16 safetensorsBF16 · flux-2-swift-mlx112.81 GBDoes not fitest.
32 GBBF16 safetensorsBF16 · flux-2-swift-mlx112.81 GBDoes not fitest.
48 GBQ4_K_M GGUFQ4_K_M · ComfyUI34.75 GBunknownunknown
64 GBQ4_K_M GGUFQ4_K_M · ComfyUI34.75 GBunknownunknown
96 GBBF16 safetensorsBF16 · flux-2-swift-mlx112.81 GBRuns · PublishedPublishedApple M2 Ultra · 96 GB RAM
128 GBBF16 safetensorsBF16 · flux-2-swift-mlx112.81 GBRuns · Published on M2 Ultra · 96 GBPublishedApple M2 Ultra · 96 GB RAM

How does it rank in blind votes?

#38 of 82 in text to image blind votes, from 77,988 votes.

Quality: blind human votes from flux-2-dev (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 M2 Ultra · 96 GB RAM1024x1024 · 28 steps · BF16 · Dev 32B, seed 42, macOS Tahoe 26.2, MLX Swift 0.30.2. Inference time only. Peak is transformer, VAE and activations, not whole-process memory.flux-2-swift-mlxSuccess1,758.6 s71 930 MBflux-2-swift-mlx quantization benchmarkPublished 2026-02-11 · 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

  • flux-2-swift-mlx

    Community

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

    Swift and MLX app and CLI for FLUX.2. It loads the BF16 transformer from the Black Forest Labs repository and takes the text encoder from separate MLX repositories. Its published peak covers transformer, VAE and activations.

    Sources: flux-2-swift-mlx README · flux-2-swift-mlx model registry

  • 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?

"Official deployment documentation says 24 to 32GB of VRAM can use 4-bit quantization. This is an NVIDIA statement, not a Mac minimum."

Source: FLUX.2 deployment guide · Read 2026-10-24.

What known issues should you check?

Frequently asked questions

Can a Mac run FLUX.2 [dev]?

One run exists on an M2 Ultra with 96 GB. flux-2-swift-mlx at BF16 took 1758.6 s for a 1024x1024 image in 28 steps, peaking at 71.93 GB for transformer, VAE and activations. Whether the time includes text encoding is not stated.

Can I use FLUX.2 [dev] commercially?

Model use is non-commercial under the FLUX Non-Commercial License. Output may be used for any purpose, including commercial, except as expressly prohibited, and may not train, fine-tune or distill a competing FLUX model.

Is there a GGUF build of FLUX.2 [dev]?

Yes. City96 publishes Q8_0 and Q4_K_M files that work in ComfyUI with the ComfyUI-GGUF custom node. The GGUF is the main model only, so both builds here pair it with a Q4_K_M GGUF Mistral text encoder and the VAE.

Sources