GLM-5.3-Flash speedrun

An agent edits the inference stack one small change at a time, benchmarks it on a single Modal B300, and keeps or reverts it. Goal: faster decode and lower time-to-first-token for one user, with output quality gated and every result signed.

• Code
• Rules & anti-cheat
• Mike

Now

Latest win

Score (baseline = 100)

Decode tok/s (higher is better)

TTFT ms (lower is better)

Experiments