The Local Coding AI Tier List — The Best Free Coding Model at Every Memory Budget, Actually Tested

The companion to the video. We put seven free, open-weights coding models through four identical tests on one 24GB MacBook Pro — a planted-bug fix, a build graded by a hidden 7-test suite, a fully agentic tool loop, and raw speed from the runtime's exact token counters. THIS SHEET: (1) THE LADDER — the tested pick at 8, 16, 24, 32, and 64GB of RAM, with the honest reasons (the value king swept every test at 45 tokens/sec from just 14GB on disk). (2) THE FULL SCORECARD — every pass, every fail, every tok/s, including the community favorite that failed the tool test and the benchmark darling that found our bug and then wrote an essay instead of fixing it. (3) THE AGENTIC DIVIDER — what tool-calling failure actually looks like from the inside, and the 2-minute check that tells you whether ANY model can act on your code or only talk about it. (4) FITS vs USABLE — the memory-ceiling trap: the same task took 36 seconds on our 14GB winner and 17 MINUTES on a 20GB heavyweight on the same Mac, and the ollama ps check that spots the spill. (5) SETUP IN 3 COMMANDS + the honest local-vs-cloud decision table. All numbers from real runs with saved transcripts, July 30, 2026. Independent sheet from Hyperautomation Labs — not affiliated with any model vendor.

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.