Tinker API - A model fine-tuning API from Thinking Machines Lab

What is the Tinker API?

Tinker API, the first product released by Thinking Machines Lab, is designed specifically for fine-tuning language models. 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 as...forward_backwardandsampleTinker enables developers to build custom fine-tuned or reinforcement learning algorithms, supporting various open-weight models from small to large, with switching models requiring only a modification to a single string in the code. Tinker integrates LoRA technology, allowing multiple training tasks to share a computing resource pool, optimizing cost-effectiveness. The open-source library Tinker Cookbook has been released, containing implementations of various post-training methods. As a managed service running on Thinking Machines' internal cluster, Tinker handles the tedious tasks of task scheduling, resource allocation, and fault recovery, allowing developers to focus on algorithms and data. Currently in free private testing, a usage-based pricing model will be launched in the future.

Main functions of the Tinker API

• Low-level operation primitives:supplyforward_backwardandsampleIt includes operation primitives that support the construction of custom fine-tuning and reinforcement learning algorithms.
• Broad model supportIt is compatible with open weight models ranging from small to large, such as Qwen-235B-A22B. Model switching only requires modifying a single string in the code.
• LoRA technology integrationThis allows multiple training tasks to share a computing resource pool, reducing costs.
• Open source support librariesReleased Tinker Cookbook, which includes implementations of various post-training methods.
• Managed servicesIt runs on the internal cluster of Thinking Machines and automatically handles task scheduling, resource allocation, and fault recovery.
• User-friendlyProvides a native Python interface, making it easy to learn and use.

Tinker API project address

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

How to use the Tinker API

• Request accessVisit the official Tinker application page: https://form.typeform.com/to/jH2xNWIg to apply to be added to the user whitelist.
• Installation and configurationInstall the Tinker API and configure the relevant environment.
• Write codeUse the low-level operation primitives provided by Tinker to write fine-tuning or reinforcement learning algorithms.
• Run trainingSubmit training tasks to Thinking Machines' managed infrastructure for execution.

Application scenarios of Tinker API

• Formal theorem proofThe Goedel team at Princeton used Tinker and LoRA to fine-tune a large language model for formal theorem proving, achieving performance comparable to a fully parameter-supervised fine-tuning model with only 20% of the data.
• Chemical reasoning model trainingRotskoff's lab at Stanford University used Tinker to train a chemical reasoning model. After reinforcement learning on top of the LLaMA 70B model, the accuracy of converting IUPAC nomenclature to chemical formulas jumped from 15% to 50%.
• Multi-agent reinforcement learningThe SkyRL team at Berkeley ran a custom multi-agent reinforcement learning loop involving asynchronous off-policy training and multi-round tooling.
• Long-context AI control tasksRedwood Research used Tinker to train the Qwen3-32B model for reinforcement learning on long-context AI control tasks.
• Classic Supervisory Fine-TuningTinker supports a variety of applications, from classic supervised fine-tuning to highly experimental reinforcement learning pipelines.