DeepSeek open-sources inference optimizations with 60–85% faster generation [pdf] Hacker News · 06-27 18:18