The Night-Shift Kit — the 24GB Local-Qwen Setup (Both Fixes Included), the Full 15-Task Results, and the Send-Local / Keep-Paid Split Sheet

The companion to the video where a free local model — Qwen3.8-27B — passed 14 of the 15 real coding tasks a $200/mo Claude Code plan passed, on an ordinary 24GB MacBook, with pass/fail scripts locked before either model ran. PAGE 2 — THE 24GB SETUP: the quant rule in one bar (56GB full model — forget it; 17GB influencer quant — suffocates 24GB; 13.1GB Q3 — the one that fits), the exact ollama pull, the 32K Modelfile rebuild (Claude Code's ~20K opening prompt cannot fit the default 16K window), and the exact env from the battery run. PAGE 3 — THE 65-LINE SHIM, VERBATIM: Qwen's chat template rejects Claude Code's mid-conversation system messages and dies with a template error; this proxy fixes it — the complete file, copy-paste ready. PAGE 4 — THE RECEIPTS: all 15 tasks with both sides' wall times (git bisect, letter-perfect file:line citations, a ~1000× perf fix, 3/3 planted review bugs found), the one honest failure (the machine RAM-cliffed three runs in a row — 24GB doesn't degrade, it cliffs), and the what-remains-unproven box. PAGE 5 — THE SPLIT SHEET: what to send to the free night shift tonight (bug-fix queue, tests, refactors, reviews, scripts from specs) vs what to keep on the paid plan (interactive pace, big one-shot generations, >32K context), plus the honest speed math: 9.4× slower overall, the viral '3× faster' setting measured at ~1.4× on a real Mac, and the $5.66 API-equivalent twist. Independent kit from Hyperautomation Labs — not affiliated with Anthropic or Alibaba.

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.