Mira Murati unveils Tinker, her AI start-up's first product
What's the story
Mira Murati's AI start-up, Thinking Machines Lab, has officially launched its first product: Tinker. It is a flexible API designed to simplify the fine-tuning of large and small open-weight AI models. Tinker aims to empower researchers and developers by abstracting the complexities of distributed training, enabling them to experiment with models and customize them to their specific needs. The service is currently in private beta, with a waitlist open for interested users.
Product features
Tinker enables advanced research and customization
Tinker is designed to enable advanced research on state-of-the-art models and their customization according to specific needs. It provides clean abstractions for writing experiments and training pipelines while handling the complexity of distributed training. The platform supports novel research, custom models, and solid baselines with minimal code changes. Murati, who previously served as OpenAI's Chief Technical Officer, announced the launch of Tinker on X, expressing her excitement about what people would build using this tool.
Platform potential
Tinker has been tested by several research groups
Tinker is a managed service that simplifies fine-tuning for both large and small open-weight AI models. The platform's flexibility is such that switching from a lightweight model to a massive mixture-of-experts system can be done with just one line of Python code. Tinker has been tested by several research groups including Princeton's Goedel Team, Stanford's Rotskoff Chemistry group, Berkeley's SkyRL group, and Redwood Research.
Developer support
Tinker Cookbook launched to support developers
Along with the launch of Tinker, the lab is also releasing the Tinker Cookbook. This open-source library comes with modern post-training methods built on the Tinker API. The aim is to help developers achieve better results while using this innovative platform. Currently, Tinker is available in private beta with a waitlist open for researchers and developers who want to try it out.