ComfyUI on a $899 M4 Mac Mini 16GB: Full Benchmark Results and Workflows
FaatmanSlim · reddit · 2026-10-11
The author benchmarked local ComfyUI workflows on a $899 M4 Mac mini (10-core CPU/GPU, 16GB unified memory), using Codex to automate the entire campaign via Python scripts, the ComfyUI API, and Comfy MCP. Running ComfyUI 0.39.2 with PyTorch 2.12.1 on Apple MPS and Q4 GGUF quantization:
- Text-to-image: FLUX.2 Klein 4B was fastest (512×512 in 40s cold / 25s warm, max 1152×1152 at 80–90s); Z-Image Turbo did 512×512 in 85s cold (max 896×896); Krea 2 Turbo 161s cold (max 768×768); Qwen Image 2.1 was slowest at 256s cold (max 896×896, 512–527s)
- Image editing: FLUX Klein single-reference editing at 512×512 in 50s cold; Qwen Image editing 271s
- SeedVR2 upscaling: 3B INT8 handled 512→1024 in 96s; FP16 slightly slower
- Qwen3-TTS: 0.6B FP32 generated 28s of audio in 60–65s; 1.7B FP16 in 85–96s
Resolutions at or above 1024×1024 generally hit a swap guard (stopped when process swap grew by 1GiB or more). The author details GGUF loader and encoder configs (Qwen3-4B/Qwen3-VL encoders) needed to get each model running on low-end Macs. Bottom line: quantized models make mainstream image generation/editing and TTS workable on 16GB, but resolution ceilings are tight.
More from Infra
- Is AI Spreading Like a Virus? Reddit Essay Questions the $700B Datacenter Buildout — thepotatobake · 2026-10-11
- Recall: Open-Source Local Semantic File Search Hits 95% Accuracy in 0.17s on an i5 Laptop — somthing_tn · 2026-10-11
- Running 426GB MXFP4 DeepSeek on 192GB VRAM to power a 4-sub-agent code review — HankYeomans · 2026-10-11
- awesome-local-ai: a task-sorted local AI tool list built for coding agents — blaizedsouza · 2026-10-11
- A single prompt made an AI find a bug eating 84% of backend CPU — zealcaiden · 2026-10-11
- Asana cut a Codex agent's cost 4x to $0.47 per run with an 89% prompt cache hit rate — daniel_mac8 · 2026-10-11