ThinkMesh: Parallel Reasoning for LLMs

ThinkMesh is a Python library for running diverse reasoning paths in parallel, scoring them with internal confidence signals, reallocating compute to promising branches, and fusing outcomes with verifiers and reducers. It supports offline Hugging Face Transformers and vLLM/TGI, and hosted APIs. ThinkMesh offers various strategies like DeepConf, Self-Consistency, and Tree of Thoughts, and includes features like caching, metrics, and JSON tracing for enhanced efficiency and reliability in large language model reasoning.