Bespoke Labs

Data and RL environments for agent training

Bespoke Labs is an applied AI research lab focused on curating data and reinforcement learning environments for training and evaluating agents. It curated Open Thoughts, among the best open reasoning datasets and used by multiple frontier labs, trained specialised models including Bespoke-MiniChart-7B and Bespoke-MiniCheck, and taught agents multi-turn tool-calling through reinforcement learning. It is also a core contributor to Terminal-bench, a leading agent benchmark used by most frontier labs — building both the training environments and the yardsticks the field measures itself against.