The Kimi K3 Swap Kit — Run the Largest Open Model Ever Inside Claude Code

The companion to the video. On July 27, 2026 Moonshot AI released the full weights of Kimi K3 — 2.8 trillion parameters, the largest open-weights model ever — and we gave it a full workday inside Claude Code, head-to-head against Claude on identical tasks. THIS KIT: (1) THE 60-SECOND SWAP — the three environment variables that point Claude Code at Kimi K3, plus the official extras most guides miss, and the [1m] model-name trick that unlocks the full 1,000,000-token context window (plain 'kimi-k3' silently caps you at 200K — we verified both). (2) THE REAL WORKDAY SCORECARD — five duels (bug hunt, refactor, feature build, hallucination trap, 825,343-token deep-context test) with times and real token costs on both sides; ten runs, ten green, whole day on Kimi: $3.36. (3) THE 51% QUESTION — what the independent hallucination number actually means, the trap we set with two invented APIs, and the practical rule for when to trust K3 (agentic with tools) vs when not to (citation-free recall). (4) THE MILLION-TOKEN DECISION — when swallowing a whole codebase beats agentic search, the cost math ($2.50 per 800K-token prompt vs $0.36 for search-first), and the auto-compact setting that makes [1m] usable. (5) THE LICENSE IN PLAIN ENGLISH — the two commercial triggers in the custom 'Kimi K3 License' ($20M MaaS revenue → separate agreement; 100M MAU or $20M/month → UI attribution) and who actually needs to care. All numbers verified July 30, 2026 from real runs with receipts. Independent kit from Hyperautomation Labs — not affiliated with Moonshot AI 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.