Apple Silicon Macs for local AI
What runs in ComfyUI on a Mac with 16, 24, 32, 48, 64 or 128 GB of unified memory, which files work on Apple GPUs, and how slow it really is.
01Pick your memory size
On a Mac what fits depends on the memory size, not the chip; the chip decides the speed. Each page lists every model with the file to download.
02How much of the memory the GPU can use
A Mac has one pool of memory for the CPU and the GPU. macOS lets the GPU use only part of it by default. On current macOS (26 and later) that is about 74 % on a 16 GB Mac, rising to about 84 % on 128 GB; older versions allowed 2/3 up to 32 GB and 3/4 above. The verdicts on this site use the current figure.
ComfyUI can use more than this, because PyTorch lets it go past the limit, but then macOS compresses and swaps memory and everything slows down sharply. You can raise the limit until the next restart with sudo sysctl iogpu.wired_limit_mb=40000 (the number is in MB; 0 restores the default). Leave several GB for macOS itself.
| Unified memory | GPU can use (macOS 26+) | Older macOS |
|---|---|---|
| 16 GB | 12.7 GB | 11.5 GB |
| 18 GB | 14.4 GB | 12.9 GB |
| 24 GB | 19.6 GB | 17.2 GB |
| 32 GB | 26.8 GB | 22.9 GB |
| 36 GB | 30.2 GB | 29 GB |
| 48 GB | 40.2 GB | 38.7 GB |
| 64 GB | 55.7 GB | 51.5 GB |
| 96 GB | 85 GB | 77.3 GB |
| 128 GB | 115.4 GB | 103.1 GB |
| 192 GB | 175.4 GB | 154.6 GB |
| 256 GB | 236.9 GB | 206.2 GB |
| 512 GB | 498.1 GB | 412.3 GB |
Apple does not publish a formula. The figures come from the “recommendedMaxWorkingSetSize” value that apps like llama.cpp and Ollama print on real Macs, converted to the same GB as the file sizes. Measured examples below.
| RAM | Measured | Share | Mac · when | Source |
|---|---|---|---|---|
| 8 GB | 5.33 GiB | 67 % | M2 MacBook Air 8GB · 2023 | github.com → |
| 16 GB | 10.67 GiB | 67 % | M1 MacBook Pro 16GB · 2023 | github.com → |
| 16 GB | 10.67 GiB | 67 % | M2 MacBook 16GB · issue dated 2025-11-02 | github.com → |
| 16 GB | 10.67 GiB | 67 % | M4 Mac mini 16GB · 2025 (date not verified) | github.com → |
| 16 GB | 11.84 GiB | 74 % | M2 Mac mini 16GB · issue dated 2026-03-18 (macOS version not stated) | github.com → |
| 24 GB | 16 GiB | 67 % | M3 MacBook Air 24GB · ~Aug 2025 | github.com → |
| 32 GB | 21.33 GiB | 67 % | M1 Max 32GB · 2023 | github.com → |
| 32 GB | 21.33 GiB | 67 % | M1 Pro 32GB · issue dated 2025-08-05 | github.com → |
| 32 GB | 24.96 GiB | 78 % | M2 Max 32GB, macOS 26.5.2 · 2026 | modelpiper.com → |
| 32 GB | 24.96 GiB | 78 % | 32GB Mac, macOS 27 · 2026 | github.com → |
| 36 GB | 27 GiB | 75 % | M3 MacBook Pro 36GB · issue dated 2024-02-06 | github.com → |
| 48 GB | 37.44 GiB | 78 % | M4 Pro MacBook Pro 48GB, macOS 26.3 · issue dated 2026-03-05 | github.com → |
| 64 GB | 48 GiB | 75 % | M1 Max MacBook Pro 64GB · 2023 | github.com → |
| 64 GB | 48 GiB | 75 % | 'M2 Mac Pro' 64GB, Darwin 23.3 (macOS 14.3) · 2024 | github.com → |
| 64 GB | 51.84 GiB | 81 % | M4 Pro Mac mini 64GB · issue dated 2026-03-01 (macOS not stated) | github.com → |
| 128 GB | 107.52 GiB | 84 % | M4 Max 128GB, macOS 26.2 (25C56) · article 2026-08-18 | ithelp.ithome.com.tw → |
| 128 GB | 107.52 GiB | 84 % | Mac Studio M4 Max 40C 128GB · ~2025 (build host darwin24.4; runtime macOS not certain) | github.com → |
| 128 GB | 107.52 GiB | 84 % | M5 Max 128GB · measured 2026-08-28 | github.com → |
| 512 GB | 464 GiB | 91 % | 512GB Mac Studio (reporter wrote 'M4'; 512GB only existed on M3 Ultra at that time) · 2026 | github.com → |
More: developer.apple.com · github.com · stencel.io · github.com
03Which file types work on a Mac
| File type | On Apple GPUs | Details |
|---|---|---|
| 16-bit (BF16 / FP16) | Works | BF16 needs macOS 14 or later and is about half the speed of FP16 on M1–M3. |
| GGUF (Q8_0 … Q2_K) | Works | Through the ComfyUI-GGUF nodes. The main way to fit big models on a Mac. |
| FP8 | No memory saving | Apple GPUs have no FP8. ComfyUI converts plain FP8 files back to 16-bit at load; “fp8_scaled” files fail with “Trying to convert Float8_e4m3fn to the MPS backend”. |
| INT8 (int8_convrot) | No memory saving | Loads in current ComfyUI by converting to higher precision, so it is not smaller in memory than 16-bit. |
| NVFP4 | No | NVIDIA RTX 50 only. |
ComfyUI runs on Apple Silicon through PyTorch’s MPS backend (ComfyUI Desktop for macOS needs macOS 13 or later). Many FP8-only community workflows need their files swapped for GGUF or 16-bit. Sources: github.com · docs.comfy.org · docs.comfy.org · raw.githubusercontent.com · github.com · github.com
04How fast is it?
Slow, compared with NVIDIA. A Mac can load models no consumer graphics card can, but each step takes several times longer: the GPU is smaller and memory bandwidth (up to about 800 GB/s on Ultra chips, 1,200 on M5 Ultra) is below a big RTX card. Image models are usable; 14B video models take many minutes to hours per clip.
| Mac | Model and setting | Result | Date | Source |
|---|---|---|---|---|
| M3 Max 128GB MacBook Pro | FLUX.1 (dev/schnell; poster tested both), bf16 weights ComfyUI (PyTorch 2.4.0) · 20 steps, resolution not stated | 2.40 s/it (20 steps in 47s; 'Prompt executed in 62.16 seconds'); a 2nd run 7.22 s/it output was blurry/noisy (known PyTorch 2.4 MPS issue at the time) | 2024-09-13 | github.com → |
| M1 Max 32GB MacBook Pro | FLUX.1 dev Q4_1 GGUF ComfyUI (torch 2.6.0.dev20240916) · 832x1216 | 22.25 s/it; ~9.5 min whole workflow | 2024-09 | medium.com → |
| M2 Mac mini 16GB (macOS 14.5) | FLUX.1 dev Q4_K_S GGUF + T5 Q4_K_M GGUF ComfyUI + ComfyUI-GGUF · 512x512 / 1024x1024, steps not stated | 512x512 ~252 s (~14.7GB used, 1.6GB swap); 1024x1024 ~760 s (4.6GB swap) | 2024-08-20 | mybyways.com → |
| Mac mini M4 Pro 48GB | FLUX (variant not stated) ComfyUI · 1024x1024 | ~300 s/image before tuning -> ~83 s/image after PYTORCH_MPS_HIGH_WATERMARK_RATIO=0.0, --force-fp16, --highvram | 2026-01-26 | note.com → |
| Mac mini M4 Pro 24GB | FLUX.1 dev Q6_K; SDXL ComfyUI (also Draw Things) · FLUX 1024x1024 20 steps; SDXL 1024x1024 25 steps | FLUX ~50-90 s/image; SDXL ~20-40 s/image; Draw Things FLUX 1024x1024 ~50 s author gives ranges, little methodology | 2026-02-15 | www.heyuan110.com → |
| M4 Max 128GB MacBook Pro | FLUX.1 schnell mflux (MLX) · 1024x1024, 2 steps | 7.50 s/it; 38.2 s total incl. model load | 2024-11-09 | github.com → |
| M3 Max / M2 Ultra / M2 Max 96GB / M1 Max 64GB / M3 Pro 36GB / M2 Pro 32GB | FLUX.1 schnell mflux (MLX), README community table · 1024x1024, 2 steps, total 'time mflux-generate' | M3 Max ~20 s (@karpathy); M2 Ultra <15 s (@awni); M2 Max 96GB ~25 s; M1 Max 64GB ~55 s; M2 Pro 32GB ~54 s; M3 Pro 36GB ~80 s; M1 Pro 16GB ~175 s | 2024 (mflux 0.4.0) | pypi.org → |
| M5 Max (RAM/GPU cores not stated) | FLUX.1 dev 4-bit mflux-based custom denoising loop · 1024x1024, 28 steps | 119.33 s median (12 runs, sd 2.89 s), uncached baseline | 2026-07-16 | dev.to → |
| M4 Max 40C 48GB; M4 Max 32C 36GB; M3 Ultra 60C 96GB | SDXL; FLUX.1 dev Q6 (M3 Ultra FP16); FLUX.1 schnell Q6 Draw Things · SDXL 1024x1024 30 steps DPM++ 2M Karras | SDXL: M4 Max 40C 15-20 s, M4 Max 32C 18-25 s, M3 Ultra 12-16 s; FLUX dev: M4 Max 40-60 s, M3 Ultra 30-45 s; schnell: M4 Max 8-12 s, M3 Ultra 4-6 s. Peak memory FLUX dev: 28GB FP16 / 16GB Q8 / 12GB Q6_K labelled 'measured' but ranges only, methodology unclear (low confidence) | 2026-07-06 | mypcrig.com → |
| Mac Studio M4 Max 40C 128GB | IllustriousXL (SDXL-based) + 5 LoRAs Stability Matrix / not clearly stated · batch 8 | ~42 s/image vs ~11 s on RTX 4090 | 2025-04-03 | forums.macrumors.com → |
| M1 Max 64GB MacBook Pro | Wan 2.2 14B T2V Q4_K_S GGUF (high+low noise) ComfyUI + ComfyUI-GGUF · 832x480, 33 frames, 20 steps (10+10) | 1 h 22 min 45 s total (250 s/step high-noise, 236 s/step low-noise, VAE ~1:41); FP8 version failed on MPS | not verified (2026) | lilting.ch → |
| M4 Pro (RAM not stated) | Wan 2.1 T2V 14B GGUF ComfyUI + GGUF loader · low resolution, ~2 s video | ~5 minutes | 2025-02-28 | huggingface.co → |
| Mac Studio M1 Max 64GB (macOS 26.3) | Qwen-Image-Edit-2511 bf16 (38GB file) ComfyUI 0.16.4, PyTorch 2.10.0 · 4 steps, resolution not stated | BF16 ~3:59; FP16 + FP32 attention upcast 2:33; FP16 alone black image; mflux Q8 + Lightning 4 steps 2:28 | 2026-03-26 | lilting.ch → |
| M1 Max 64GB MacBook | Qwen-Image-Edit-2509 GGUF (Q8_0/Q4_K_S) ComfyUI + GGUF (vs stable-diffusion.cpp) · two input images, 512x512 in sd.cpp test | ComfyUI ~3-4 min per edit; sd.cpp Metal 20 steps ~9 min | 2025-12-02 | soywiz.com → |
reported Other people’s numbers, copied as published. Settings and software differ a lot; several were run in other apps (mflux, Draw Things), which are often faster than ComfyUI on a Mac.
05Every Apple Silicon chip
| Chip | Year | Memory options | GB/s | Macs | Specs |
|---|---|---|---|---|---|
| M1 | 2020 | 8 · 16 GB | 68.25 | MacBook Air (M1), MacBook Pro 13 (M1), Mac mini (M1), iMac 24 (M1, 2021) | apple.com → |
| M1 Pro | 2021 | 16 · 32 GB | 200 | MacBook Pro 14/16 (2021) | apple.com → |
| M1 Max | 2021 | 32 · 64 GB | 400 | MacBook Pro 14/16 (2021), Mac Studio (2022) | apple.com → |
| M1 Ultra | 2022 | 64 · 128 GB | 800 | Mac Studio (2022) | apple.com → |
| M2 | 2022 | 8 · 16 · 24 GB | 100 | MacBook Air 13 (M2, 2022), MacBook Air 15 (M2, 2023), MacBook Pro 13 (M2, 2022), Mac mini (2023) | apple.com → |
| M2 Pro | 2023 | 16 · 32 GB | 200 | MacBook Pro 14/16 (2023), Mac mini (2023) | apple.com → |
| M2 Max | 2023 | 32 · 64 · 96 GB | 400 | MacBook Pro 14/16 (2023), Mac Studio (2023) | apple.com → |
| M2 Ultra | 2023 | 64 · 128 · 192 GB | 800 | Mac Studio (2023), Mac Pro (2023) | apple.com → |
| M3 | 2023 | 8 · 16 · 24 GB | 100 | MacBook Pro 14 (M3, Nov 2023), iMac 24 (M3, 2023), MacBook Air 13/15 (M3, 2024) | apple.com → |
| M3 Pro | 2023 | 18 · 36 GB | 150 | MacBook Pro 14/16 (Nov 2023) | apple.com → |
| M3 Max | 2023 | 36 · 48 · 64 · 96 · 128 GB | 300 / 400 | MacBook Pro 14/16 (Nov 2023) | apple.com → |
| M3 Ultra | 2025 | 96 · 256 · 512 GB | 819 | Mac Studio (2025) | apple.com → |
| M4 | 2024 | 16 · 24 · 32 GB | 120 | MacBook Pro 14 (M4, 2024), Mac mini (2024), iMac 24 (M4, 2024), MacBook Air 13/15 (M4, 2025) | apple.com → |
| M4 Pro | 2024 | 24 · 48 · 64 GB | 273 | MacBook Pro 14/16 (2024), Mac mini (2024) | apple.com → |
| M4 Max | 2024 | 36 · 48 · 64 · 128 GB | 410 / 546 | MacBook Pro 14/16 (2024), Mac Studio (2025) | apple.com → |
| M5 | 2025 | 16 · 24 · 32 GB | 153 | MacBook Pro 14 (M5, Oct 2025), MacBook Air 13/15 (M5, Mar 2026) | apple.com → |
| M5 Pro | 2026 | 24 · 48 · 64 GB | 307 | MacBook Pro 14/16 (Mar 2026), Mac mini (2026, available Sept 22 2026) | apple.com → |
| M5 Max | 2026 | 36 · 48 · 64 · 128 GB | 460 / 614 | MacBook Pro 14/16 (Mar 2026), Mac Studio (2026, available Sept 22 2026) | apple.com → |
| M5 Ultra | 2026 | 96 · 256 · 512 GB | 1200 | Mac Studio (2026; pre-order Aug 25, available Sept 22 2026) | apple.com → |
| M6 | 2026 | 16 · 24 · 32 GB | 153 / 170 | Mac mini (2026, available Sept 22 2026) | apple.com → |
Memory bandwidth from Apple’s specs (M1’s 68 GB/s is a third-party figure; Apple did not publish one). Where two figures are given, they are for the two GPU core counts.