Toward Open Weight Models Without Risks: Separating Public and Private Capabilities in LLMs
Tiered Language Models (TLMs) provide a framework for releasing large language models with configurable capability levels through secret keys that modify computation graphs while m…