Mira Murati's Thinking Machines Launches Tinker API to Democratise Model Fine-Tuning

Ex-OpenAI CTO Mira Murati's startup Thinking Machines Lab today unveiled Tinker, a flexible API designed to simplify fine-tuning of large language models (LLMs).

The offering is now in private beta, targeting researchers, developers, and institutions seeking more control over model training without the complexity of managing infrastructure.

Tinker allows users to fine-tune a variety of open-weight models — from smaller ones to large mixture-of-experts models like Qwen-235B-A22B — by simply changing a string in Python code.

"Tinker is a managed service that runs on our internal clusters and training infrastructure. We handle scheduling, resource allocation, and failure recovery. This allows you to get small or large runs started immediately, without worrying about managing infrastructure," the startup said.

The platform is built on LoRA (Low-Rank Adaptation) techniques, which enable efficient training by modifying only a small part of a model instead of retraining entire weights. This approach allows multiple users to share compute resources and lowers cost.

Tinker also includes the Tinker Cookbook, an open-source library with ready-made implementations of common training and post-training routines, aiming to accelerate development.

Already, teams at Princeton, Stanford, Berkeley, and Redwood Research have used Tinker for specialized tasks — from theorem proving to multi-agent reinforcement learning loops.

“It empowers researchers and hackers to experiment with models by giving them control over the algorithms and data while we handle the complexity of distributed training,” the startup added.

Access to Tinker is free during the beta period, with usage-based pricing expected in the coming weeks. Developers and organizations can sign up via the waitlist.