Nick Trenkler
hermes is now officially faster than codex. here's how they did it
hermes was losing to codex 6–5 across 11 benchmark tasks measuring real-world cli speed – things like file editing, codebase exploration, shell pipelines, and multi-turn sessions
teknium shipped 3 fixes:
• bitwarden disk cache – the app was calling a password manager api on every single startup. added a local cache so it only fetches once
• lazy model loading – a huge list of ai providers was loading into memory at startup even when not needed. now it only loads when actually used
• config file dedupe – the same config file was being read twice on startup for no reason. merged into one read
total: 701ms → 258ms. that's -63% startup time. hermes now wins 6–5. the scoreboard just flipped
nous research vs openai. the war has just started
Some new improvements to performance just went in.
— Teknium 🪽 (@Teknium) May 25, 2026
Python gets a bad wrap for performance but we aint looking to shabby against a trillion dollar co's rust codebase, beating codex at most multi-turn tasks we benchmarked (mt stands for multiturn)
PR: https://t.co/nNJz3EPAPA pic.twitter.com/6z7vXpLA9X
