Echo is an AI system that utilizes a pool of open-weight models, rather than relying on a single model for all tasks. The project began with an experiment where multiple models were run on the same evaluations. A hypothetical system that could intelligently combine their outputs performed significantly better than any individual model.

Echo aims to replicate this advantage without prior knowledge. For each request, it dynamically allocates computation, selects participating models, and orchestrates their combined output. The developer found that even seemingly weaker models can be highly complementary and useful in specific scenarios or as part of a combination.

In initial evaluations, Echo consistently outperformed the best individual model in its pool. It achieved comparable results to Fable at approximately one-third the inference cost. While challenges remain in optimal allocation and combination decisions, the developer is actively analyzing failures and testing Echo on coding and agentic tasks. A chat interface and an OpenAI-compatible API are available for testing.