The 2GB Local AI Engine — Run a 26B-Parameter Model on the Mac You Already Own, Fully Offline, in Under 2GB of RAM

The companion to the video. turbo-fieldfare (github.com/drumih/turbo-fieldfare, Apache-2.0, 4,863 GitHub stars as of Aug 4, 2026) runs Gemma 4 26B-A4B IT in 4-bit on Apple Silicon by keeping only the shared 1.35GB core plus the conversation memory in RAM and streaming the routed experts from your SSD, with the 16 hottest experts cached — so the whole engine runs in under 2GB of resident memory. INSIDE: (1) THE REQUIREMENTS CHECKLIST — Apple Silicon only, macOS 26 with Metal 4, Swift 6.2+ command-line tools (no full Xcode — we verified), ~30GB free disk during install, ~14.3GB after; confirm every row before you clone anything. (2) THE EXACT INSTALL — the four commands start to finish (build took our M4 Pro 88 seconds), plus the --resume / --discard-partial flags and when HF_TOKEN matters. (3) THE THREE DOORS — the Mac app, the CLI with its stderr timing footer, and the OpenAI-compatible local server, including how to point any OpenAI-compatible tool at your own Mac. (4) THE PER-CHIP SPEED TABLE WITH ATTRIBUTION — our verified M4 Pro numbers (13-16 tok/s short/medium, 9.2 on the longest run, 1.53-1.57GB peak RSS) alongside the author's M2 Air 8GB (5-6) and M5 Pro (31-35) claims and an HN commenter's M4 Max 48 tok/s at 1.9GB — every row labeled with who measured it. (5) TROUBLESHOOTING — the 3 real errors we hit and their fixes. (6) THE HONEST CATCHES — one pinned text-only model, macOS 26 only, and the failed self-written unit test that shows why you verify what it drafts. Every receipt from tests run on this machine Aug 4, 2026. Independent guide from Hyperautomation Labs — not affiliated with the project or Google.

By submitting, you'll be subscribed to the Hyperautomation AI Report newsletter — one practical AI email per week. Unsubscribe anytime.

Free. No spam. Unsubscribe anytime.