Tinker API - A model fine-tuning API launched by Thinking Machines Lab

What is the Tinker API

The Tinker API is the first product released by Thinking Machines Lab, designed specifically for language model fine-tuning. It simplifies the language model fine-tuning process, allowing researchers and developers to focus on algorithms and data without worrying about complex distributed training infrastructure. It provides low-level operation primitives, such asforward_backwardandsample, enabling developers to build custom fine-tuning or reinforcement learning algorithms, supporting various open-weight models from small to large, with switching models requiring only changing one string in the code. Tinker integrates LoRA technology, allowing multiple training tasks to share a compute resource pool and optimizing cost-effectiveness. It has released the open-source library Tinker Cookbook, which includes implementations of various post-training methods. As a managed service, Tinker runs on Thinking Machines' internal cluster, handling tedious tasks such as task scheduling, resource allocation, and failure recovery for users, allowing developers to focus on algorithms and data. It is currently in a free private beta phase, and a usage-based pricing model will be introduced in the future.

Main features of the Tinker API

• Low-level operation primitives: Providesforward_backwardandsampleand other operation primitives, supporting the construction of custom fine-tuning and reinforcement learning algorithms.
• Broad model support: Compatible with open-weight models from small to large, such as Qwen-235B-A22B, with switching models requiring only changing one string in the code.
• LoRA technology integration: Allows multiple training tasks to share a compute resource pool, reducing costs.
• Open-source companion library: Released Tinker Cookbook, which includes implementations of various post-training methods.
• Managed service: Runs on Thinking Machines' internal cluster, automatically handling task scheduling, resource allocation, and failure recovery.
• User-friendly: Provides a Python-native interface, easy to get started with and use.

Project address of the Tinker API

• Project official website：https://thinkingmachines.ai/blog/announcing-tinker/

How to use the Tinker API

• Apply for access: Visit the Tinker official application page:https://form.typeform.com/to/jH2xNWIg申请加入用户白名单。
• Installation and configuration: Install the Tinker API and configure the relevant environment.
• Write code: Use the low-level operation primitives provided by Tinker to write fine-tuning or reinforcement learning algorithms.
• Run training: Submit training tasks to run on Thinking Machines' managed infrastructure.

Application scenarios of the Tinker API

• Formal theorem proving: Princeton's Goedel team used Tinker and LoRA to fine-tune a large language model for formal theorem proving, achieving performance comparable to a full-parameter supervised fine-tuning model with only 20% of the data.
• Chemical reasoning model training: Stanford University's Rotskoff lab used Tinker to train a chemical reasoning model. After reinforcement learning on the LLaMA 70B model, the accuracy of converting from IUPAC names to chemical formulas jumped from 15% to 50%.
• Multi-agent reinforcement learning: Berkeley's SkyRL team ran a custom multi-agent reinforcement learning loop, involving asynchronous off-policy training and multi-turn tool use.
• Long-context AI control tasks: Redwood Research used Tinker to perform reinforcement learning training on the Qwen3-32B model for long-context AI control tasks.
• Classic supervised fine-tuning: Tinker supports a variety of application scenarios, from classic supervised fine-tuning to highly experimental reinforcement learning pipelines.