This page may contain stale information. Last updated: 2026-06-12
Definition
NanoGPT Speedrun is a community benchmark measuring how fast a small language model can be trained to a target validation loss (3.28). The baseline has been refined by humans (often with AI assistance) over 2+ years.
June 2026 Results
| System | Time to 3.28 val loss |
|---|---|
| Community best | 79.7s |
| recursive automated system | 77.5s |
Despite years of community optimization, Recursive’s automated research system found an additional 2.2s speedup (2026-06-11-recursive-automated-ai-research).
Improvement is modest on a heavily optimized baseline — context matters when reporting speedrun gains.