Thinking Machines

Thinking Machines Lab builds AI models and tools for customizing them. Its Tinker training API lets researchers and developers fine-tune models with their own data and algorithms while the company manages distributed GPU infrastructure. Users can run supervised learning or reinforcement learning, sample outputs, and download saved checkpoints. Its Inkling model family adds a multimodal foundation for customization: Inkling processes text, images, and audio, offers adjustable reasoning effort, and is available with full weights and fine-tuning through Tinker.

Founded in 2025, the company is led by cofounder and CEO Mira Murati, formerly OpenAI’s chief technology officer. Its research includes work on on-policy distillation, where a teacher model provides token-level feedback on a student model’s own generated sequences. Building on earlier research, the lab demonstrated how this approach can restore instruction-following behavior after a model learns internal company information, with implementations available in the Tinker cookbook.

Tinker charges by token usage and checkpoint storage, serving teams that want to control training without managing compute clusters. Glean used it to train Waldo, a search model using production traces and reinforcement learning; Chroma trained Context-1, a specialized search subagent. These applications put model customization inside document retrieval workflows rather than limiting it to standalone research experiments.