NanoGPT speedrun record falls to 39.9 seconds on 8×H100, biggest single leap yet

QuintinPope5 · x · 2026-09-29

Hyperstition trains NanoGPT (124M) to target validation loss 3.28 on FineWeb in 39.914 seconds on 8×H100, down from 73.889s — a percentage improvement exceeding the previous 45 world records combined. Gains come from ANVIL II (an optimizer beating Muon), sampled softmax, and a larger sparse n-gram table; ablations attribute 8.41s to sampled softmax, 7.06s to the n-gram table, 5.09s to ANVIL+averaging. Frontier-scaling methods were redacted. Quintin Pope notes nanochat/modded-nanogpt progress has stalled for months and the author didn't use autoresearch — efficiency headroom remains, but AI isn't good at finding it.

Related event: NanoGPT speedrun record slashed to 39.9 seconds(4 posts)→

Original post →

More from Infra

Infra channel →