Home › Research Papers › Training Compute-Optimal LLMs — Chinchilla (2022)
Training Compute-Optimal LLMs — Chinchilla (2022)
Research Paper
Scaling
DeepMind's correction: most models were undertrained on data.
You're looking at a research-paper record: a landmark paper and what it introduced to the field.
Hoffmann et al. showed compute-optimal training wants ~20 tokens per parameter — far more data than the field was using — so a 70B model trained right beat a 280B one. It reshaped every subsequent training run and made data supply a first-order constraint.
How this connects to the rest of the stack
This record's links at a glance — click a group to jump to it.
Authored by
Basis for
Related
Sources
Record reviewed as of 2026-01-15 — see Methodology.