On Aug 31, 2026 we took the free local model that had already passed 14/15 graded coding tasks inside Claude Code — a 27B Qwen running on an ordinary 24GB MacBook — and swapped Claude Code itself out for DeepSeek Harness (dsh), the MIT-licensed open-source agent harness DeepSeek released in August. Then we re-ran the identical 15-task battery — failing tests, a feature build, a git bisect, a merge conflict, strict TypeScript with cheating banned, a security hole, a performance bug, a planted-bug code review — graded by the same pass/fail scripts, locked before any model ran. Result: 14/15 again, the same fourteen, the same single miss, 19 minutes faster overall. Zero subscription, zero API key, nothing leaving the Mac. This kit is the whole setup: PAGE 2 — the 10-minute path plus the exact 30-line config file from the passing run, with the boot trap (dsh's permission-preset refusal) already fixed. PAGE 3 — the full harness-vs-harness ledger, wins and losses both ways, plus the honest paid control (Opus: 15/15 in 17 minutes — still ~8× faster). PAGE 4 — all 15 task prompts verbatim so you can grade any local model yourself. PAGE 5 — the run-local vs keep-paid decision sheet, plus the task-5 double autopsy (in Claude Code it died by memory wall; in dsh it 'thought' for 32 minutes — 11,399 reasoning chunks, zero tool calls — and never wrote a byte). Every number is a committed receipt from our own runs. Independent guide from Hyperautomation Labs — not affiliated with DeepSeek, Alibaba, or Anthropic.

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.