Former OpenAI CTO Mira Murati’s startup, Thinking Machines Lab, has revealed its first product: Tinker, an API for fine-tuning AI models.
Tinker is now in private beta; interested organizations can contact Thinking Machines Lab. The API is free as of its announcement on Oct. 1, but it will switch to usage-based pricing in “the coming weeks,” according to the blog post.
What does Tinker do?
The Tinker API enables AI and security researchers to fine-tune large language models, allowing them to experiment with the algorithms and data using Python without incurring the burden of distributed training. Models that can be accessed inside Tinker include Alibaba’s Qwen-235B-A22B and Meta’s Llama-3.2-1B, among others.
“Tinker advances our mission of enabling more people to do research on cutting-edge models and customize them to their needs,” Thinking Machines Lab stated in its announcement blog post.
Using low-rank allocation (LoRa) lowers the costs of the service because it allows Thinking Machines Lab to use the same pool of compute for multiple training runs, the company said.
Teams at Princeton University, Stanford University, University of California, Berkeley, and the nonprofit AI security organization Redwood Research have already completed projects using Tinker.
Tinker Cookbook offers suggested abstractions
The Tinker Cookbook, released alongside the Tinker API, is an open-source library that provides examples of how to use Tinker and offers shortcuts. It includes common abstractions to help researchers achieve their goals by customizing their Tinker training environments. For example, it can walk through how to fine-tune math reasoning, conversational datasets, use of retrieval tools, or prompt distillation.
Thinking Machines Lab’s founders are seasoned AI researchers
Murati founded Thinking Machines Lab in February 2025. She was joined by other veterans of the AI boom, including OpenAI co-founder John Schulman and former employees of Google, Meta, Mistral, and Character AI.
In September, Thinking Machines Lab announced its researchers had published a research paper on reducing nondeterminism, or the tendency of generative AI to provide different answers to the same question.