For the past few years, DeepMind’s Chinchilla scaling law has been the unofficial gospel of large language model training. It told researchers how to balance model size against training data to get the best performance per unit of compute. Meta’s FAIR lab just published a paper arguing that Chinchilla gets the relationship fundamentally wrong, and that fixing it could slash the compute needed to plan model training by roughly 10x.
The paper, titled “Skaling: Chinchilla’s Exponents Meet Kaplan’s Coupling,” was published on August 7, 2026 by Mathurin Videau, Badr Youbi-Idrissi, David Lopez-Paz, and Kartik Ahuja. It introduces a modified scaling law that accounts for something Chinchilla quietly ignores: the interaction between model size and training data volume.
What Chinchilla gets wrong
DeepMind’s Chinchilla scaling law, established in March 2022, treats model parameters (N) and training tokens (D) as independent variables that each contribute separately to a model’s final loss.
The Skaling paper demonstrates that this independence assumption breaks down in practice. By computing the mixed partial derivative of the loss surface with respect to N and D, the researchers showed it’s non-zero. In plainer terms, the effect of adding more parameters depends on how much data you’re training on, and vice versa.
Skaling captures this interaction with a coupling exponent, denoted k, wrapped around the standard Chinchilla formulation. The full model takes the form L(N, D) = (A / N^α + B / D^β)^k + E, where k greater than 1 represents the coupling between model size and data volume. When k equals 1, the formula collapses back to the original Chinchilla form.
The numbers make it hard to argue
Skaling outperformed Chinchilla at 76% of tested configurations, with a median improvement factor of 2.2x in prediction accuracy. At one-third of the tested points, improvements reached 4x or higher. Overall, the new law reduced mean absolute percentage error by 1.5 to 3x compared to Chinchilla across various test scenarios.
Skaling achieves nearly identical accuracy using a sparse “L-shaped” profiling grid, where you vary one dimension while holding the other fixed, then repeat in the other direction. This approach consumes approximately 10x less compute than the full-grid method Chinchilla needs.
At frontier compute levels, the optimal token-to-parameter ratio can differ by orders of magnitude from what Chinchilla would suggest.
Why this matters for model training strategy
The Chinchilla paper was transformative when it arrived in 2022. It argued that many existing models were significantly undertrained, meaning researchers were building models with too many parameters and not feeding them enough data.
Many production models are deliberately “overtrained,” meaning they see far more tokens than Chinchilla’s optimal ratio would suggest, because inference costs favor smaller models that punch above their weight. The Skaling law’s coupling exponent captures how the optimal balance shifts in these overtraining regimes. The practical upshot is that Skaling can tell you more accurately how to adjust model parameters when data availability is constrained, or when you’re deliberately overtraining a smaller model for deployment efficiency.