Hacker News
A Week of Real Kimi K3 Use: Opus-Level Instruction Following, but Overload Errors and Empty Thinking-Token Stalls
Chen Chen's July 24, 2026 field report — surfaced on HN ahead of the July 27 weights drop — finds K3 holds architectural patterns and naming conventions across dozens of files without drifting, matching Opus on instruction adherence and Fable on statistical analysis. The failure modes are operational, not capability: overload errors interrupt long sessions and require retry logic, the model sometimes emits only thinking tokens and stops with no content, and tool calling can loop burning tokens without progress. Cache hit rates depend on careful key management, and subscription credits drain noticeably faster than Claude or Codex — useful priors for anyone planning to route agent traffic to K3 next week.
↳ Follow the thread