Which Image Models Fit Your Mac's RAM?
On Apple Silicon, unified memory plays the role VRAM plays on a GPU — the model must fit in RAM alongside macOS (budget ~5GB headroom, our estimate). The file sizes below were verified from each model's HuggingFace repository on 2026-06-12.
| Mac RAM | Fits comfortably | Borderline | Recommended app |
|---|---|---|---|
| 8GB | SD 1.5 (4.27GB) | SDXL (6.94GB) — expect swapping; FLUX not practical | Draw Things or Mochi Diffusion (Core ML / Neural Engine) |
| 16GB | SD 1.5, SDXL, SD 3.5 Medium (5.11GB + 4.89GB fp8 text encoder) | FLUX.1 4-bit (MFLUX pipeline ~9.85GB) and FLUX.2 klein 4B — work, little headroom | Draw Things; MFLUX 4-bit for FLUX |
| 32GB | FLUX.1 dev GGUF Q8 (12.7GB), FLUX.2 klein 4B (7.75GB), SD 3.5 Large (16.5GB), Qwen-Image Q4 (12.1GB) | FLUX.2 klein 9B (18.2GB weights), Qwen-Image Q8 (21.8GB) | Draw Things for speed; ComfyUI Desktop for GGUF workflows |
| 64GB+ | FLUX.1 bf16 (23.8GB), Qwen-Image Q8 (21.8GB), FLUX.2-dev GGUF Q4-Q6 (19.3-27.4GB) | FLUX.2-dev bf16 (64.4GB) does NOT fit — needs 96-128GB | ComfyUI (GGUF) or MFLUX (native MLX) |
File-size sources: SD 1.5, SDXL, SD 3.5 Medium, SD 3.5 Large, FLUX.1, FLUX.1 GGUF quants, FLUX.2 klein 4B, FLUX.2-dev GGUF, Qwen-Image GGUF.
Which App Should You Use?
Draw Things — best for most people
Free · App StoreA custom native Metal engine (not a Python wrapper), one-click model downloads, and support for everything from SD 1.5 to FLUX.2, Qwen Image and LoRA training on-device. Local generation is unlimited and free; the optional subscription only covers cloud compute. drawthings.ai
ComfyUI Desktop — best for power users
Free · open source · Apple Silicon betaThe node-based standard now ships a native Apple Silicon desktop app (beta). Runs GGUF-quantized FLUX.1 and FLUX.2 via the ComfyUI-GGUF node — the key to fitting big models in limited RAM. A community benchmark on an M4 Pro Mac mini found it roughly 20% slower than Draw Things on identical hardware (heyuan110.com). comfy.org
MFLUX — best native-MLX option (CLI)
Free · open sourceA line-by-line MLX port of FLUX with built-in 4-bit/8-bit quantization and a --low-ram flag. Tracks new releases fast: FLUX.2 klein, Z-Image, Qwen Image and more. Per its own benchmark table, FLUX.1 schnell generates a 1024×1024 image in under 15 seconds on an M2 Ultra and about 30 seconds on an M3 Pro with the 4-bit pipeline (mflux README).
Mochi Diffusion — best Neural Engine efficiency
Free · open sourceRuns Core ML models on the Neural Engine — the app itself uses ~150MB of RAM per its README. Back under active development (v6.0, Feb 2026) with FLUX.2 Klein support. GitHub
Skip DiffusionBee. It was the classic beginner pick, but it has not shipped a release since August 2024 and lags badly on modern models. Most "Stable Diffusion on Mac" listicles still recommend it — that advice is two years stale.
FLUX.2 on Mac: What Runs in 2026?
FLUX.2 is open-weight, and the practical Mac question is which variant fits. FLUX.2 klein 4B (Apache 2.0, released January 2026) is the sweet spot: 7.75GB of weights, and Black Forest Labs states it fits in about 13GB of memory — workable on a 16GB Mac, comfortable on 32GB. The 9B klein (18.2GB weights, ~29GB per BFL) wants 32GB+. The full FLUX.2 [dev] is a 32B model whose bf16 checkpoint is 64.4GB — on Macs that means GGUF quants (Q4 at 19.3GB, Q6 at 27.4GB) on a 64GB machine, plus its large text encoder.
Sizes from the official FLUX.2-klein-4B, FLUX.2-klein-9B and FLUX.2-dev repos, GGUF quants from city96.
How Fast Is Image Generation on Apple Silicon?
All figures below are community-published measurements, attributed to their source — ModelFit does not run its own benchmarks:
- →SD 1.5, 512×512: ~5-10s on an M4 Pro Mac mini (per heyuan110); ~8-15s on M2 Pro in Draw Things (per rentamac.io)
- →SDXL, 1024×1024: ~20-40s on M4 Pro, 25 steps (per heyuan110)
- →FLUX.1 dev Q6, 1024×1024, 20 steps: ~50-90s on M4 Pro in ComfyUI (per heyuan110)
- →FLUX.1 schnell (2 steps): under 15s on M2 Ultra, ~80s on M3 Pro non-quantized (per the mflux benchmark table)
The pattern: schnell-class models are interactive on any recent Pro/Max chip, dev-class models are a coffee-sip wait, and base-M-chip 8GB machines should stay on SD 1.5.
Frequently Asked Questions
Can I run FLUX on a Mac?
Yes — via Draw Things, ComfyUI or MFLUX. On 16GB use 4-bit quantization (~9.85GB pipeline); 32GB runs Q8; full bf16 (23.8GB) wants 64GB. FLUX.2 klein 4B is the newer efficient option at 7.75GB.
What is the best app for local image generation on Mac?
Draw Things for most users (free, fast Metal engine, every major model). ComfyUI Desktop for node workflows and GGUF quants. MFLUX if you prefer a native-MLX command line. Avoid DiffusionBee — unmaintained since 2024.
How much RAM do I need for Stable Diffusion or FLUX on a Mac?
8GB: SD 1.5. 16GB: SDXL and quantized FLUX. 32GB: FLUX.1 Q8, FLUX.2 klein, SD 3.5 Large. 64GB: full-precision FLUX.1 and 20B-class models like Qwen-Image.
Is local image generation on Mac as good as Midjourney or DALL-E?
FLUX-class open models produce results competitive with cloud services for most use cases, with no subscription, no content filter surprises, and full privacy. The trade-offs are generation speed (seconds to a minute-plus locally) and needing enough RAM for the bigger models.