Best Local LLMs for Every Mac: Run a 27B Model on Just 16GB RAM
JosephJacks_ · x · 2026-08-06
A practical guide to choosing the best local LLM based on a Mac's unified memory. The author categorized Macs into six tiers (from 8GB to 96GB+) and tested the optimal model for each on a single M3 Ultra, recording peak RAM and boot times.
- 8–15 GB: lfm2.5-2.6b-4bit (Peak RSS: 2.0 GB).
- 16–23 GB: bonsai-27b-2bit (Peak RSS: 8.4 GB, proving 27B models can run on 16GB).
- 24–31 GB: gemma-4-26b-4bit.
- 32–63 GB: qwen3.6-35b-4bit.
- 64–95 GB: qwen3.6-35b-8bit.
- 96 GB+: qwen3.5-122b-mxfp4.
Users can automatically detect their hardware and install the recommended model using the provided curl | bash script.
More from Infra
- Google Cloud's Filestore Migrates to Colossus, Decoupling Capacity from IOPS — rseroter · 2026-08-06
- Testing 8x DGX Spark Nodes in Open World Multi-Agent Setup — NVIDIAAI · 2026-08-06
- Open-Source Benchmarks: RTX 5090 LLM Quants and 8GB VRAM Agentic Scores — max_paperclips · 2026-08-06
- NVIDIA Discusses Building Secure Enterprise AI with Proprietary Data — nvidia · 2026-08-06
- Chorus: Open-Source Pre-trained Model Library Enables Fast CPU Inference Without GPUs — jmschreiber91 · 2026-08-06
- Running DeepSeek V4 Locally on Spark Hardware Hits ~95 tok/s — Rasmic · 2026-08-06