Training a 3.8B LLM to 0.384 CORE for $998
A solo project trained a 3.8B decoder-only model on 65B tokens in 43 hours for $998 on rented B200s, reaching a reported 0.384 CORE score. The config-driven training framework makes experiments reproducible through YAML and registries, and the write-up includes failed runs, architecture choices, data, optimizer schedules, and cost trade-offs—not just a leaderboard number. Treat the comparison as one experiment, not a general scaling law, but it is a useful demonstration of what a focused individual can now investigate.
The HN thread had one comment asking how much additional budget would be needed to perform well on AIME.