NeoSmith

NeoSmith provides a framework for assigning a dedicated Small Language Model (SLM) to each component of an agentic workflow. It captures traffic from a larger language model, runs an automated distillation process, and produces a lightweight model that mirrors the behavior of the original for that specific agent. The resulting SLMs are typically under one billion parameters and claim to reduce inference cost by an order of magnitude.

The system includes pipelines that automate the distillation, tools for integrating the generated SLMs into existing workflows, and mechanisms for deploying the models in production environments. It targets developers building complex, multi‑agent AI systems who need to control cost and latency while preserving the specialized behavior of each agent.

NeoSmith is positioned as an experimental solution that emphasizes custom, per‑agent model generation rather than a single shared model, aiming to make large‑scale agentic applications more economical and easier to manage.