ModelRefs / Chinchilla Scaling — AI Glossary
Chinchilla Scaling — AI Glossary
DeepMind's 2022 finding that optimal training uses ~20 tokens per parameter, shifting the field toward smaller, better-trained models.
Overview
Hoffmann et al. showed that Gopher (280B params, 300B tokens) was undertrained: equal compute spent on 70B params trained on 1.4T tokens ('Chinchilla') matched or exceeded Gopher on nearly all tasks. This revised OpenAI's 2020 scaling laws and prompted Llama, Mistral, and Phi series designs.
Reference details
| Topic | architecture |
|---|---|
| Also known as | Chinchilla optimal, compute-optimal |
| Last reviewed | 2026-06-24 |
Related terms
Continue your research
Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Chinchilla Scaling — AI Glossary.
Frequently asked questions
What is Chinchilla Scaling?
DeepMind's 2022 finding that optimal training uses ~20 tokens per parameter, shifting the field toward smaller, better-trained models.
Is Chinchilla Scaling the same as Chinchilla optimal?
Yes — Chinchilla optimal, compute-optimal are common aliases for Chinchilla Scaling.
What concepts are related to Chinchilla Scaling?
Closely related concepts include scaling law, neural scaling, pretraining data.