DeepSeek R1 (2025) › The Problem

One approach to improving reasoning was to increase both model size and training data. This strategy improved performance on many benchmarks, including tasks that had been difficult for earlier systems. It also made parameter count and training compute major design constraints.

Training and deploying very large models required computational resources unavailable to many research groups. This limited who could develop or operate high-performing reasoning systems and made efficiency a practical concern rather than only a benchmark metric.

Model size and reasoning performance did not improve at the same rate on every task. Diminishing returns on some benchmarks suggested that adding parameters was not always the most efficient intervention. Serving cost also constrained where a reasoning system could be deployed.

Consider a scenario where a research team wanted to develop a reasoning system for deployment in resource-constrained environments, such as edge computing devices or mobile applications. Traditional scaling approaches would suggest training a model with as many parameters as possible, requiring substantial computational resources during training and significant infrastructure during deployment. However, this approach might not be feasible for applications with strict computational or energy constraints.

Scale could also obscure the contribution of individual design choices. When several dimensions increased together, researchers had difficulty distinguishing architecture effects from changes in the data or training procedure. More efficient models required clearer evidence about which changes produced a reasoning gain.

Treating scale as the only route to better reasoning would keep development and deployment tied to large computational budgets. R1 instead emphasized architecture and task-specific training as additional variables.