The power to customize advanced language models without wrestling with infrastructure or distributed training is what Tinker by Thinking Machines promises. This innovative API puts flexible, powerful fine-tuning capabilities in the hands of both experts and newcomers, eliminating traditional barriers to entry.
Seamless Model Fine-Tuning
Tinker acts as a managed service that enables users to fine-tune a wide selection of open-weight language models, from compact configurations to expansive mixture-of-expert systems like Qwen-235B-A22B. Transitioning between models is as simple as updating a single string in your code. Tinker automates the backend handling scheduling, resource management, and recovery so users can fully focus on experimentation and innovation.
Efficiency Built In
At its core, Tinker leverages Low-Rank Adaptation (LoRA) to maximize resource sharing among training runs. This approach keeps costs manageable while boosting computational efficiency. The API also provides advanced primitives such as forward_backward
and sample
, giving users granular control to implement and experiment with a variety of post-training strategies.
Accessible Advanced Techniques
While Tinker’s flexibility is a game changer, mastering language model fine-tuning still demands technical skill. To support users, Thinking Machines released the open-source Tinker Cookbook. This library offers ready-to-use implementations of post-training methods, all running seamlessly atop the Tinker API. It lets researchers and developers skip boilerplate code and focus on novel ideas.
Real-World Impact and Adoption
Tinker is already powering projects at leading research institutions:
- The Princeton Goedel Team trains mathematical theorem provers with Tinker.
- Stanford’s Rotskoff Chemistry group tackles advanced chemistry reasoning tasks.
- The SkyRL group at Berkeley explores complex reinforcement learning experiments.
- Redwood Research fine-tunes Qwen3-32B for challenging AI control problems.
How to Get Started
Tinker is currently in private beta. Interested developers and organizations can join the waitlist for early access or contact the team to discuss customized solutions. The service is free to start, with usage-based pricing planned for the future.
Unlocking AI Innovation for All
Tinker signals a new era in language model customization. By removing infrastructure headaches and offering robust, user-friendly tools, Thinking Machines is enabling more innovators to push the boundaries of AI. Whether you’re fine-tuning for a specialized application or pioneering new research methods, Tinker brings advanced model customization within everyone’s reach.
About Thinking Machines
Thinking Machines Lab was founded by the scientists, engineers, and builders who created some of the most widely used AI products and open-source projects in the world, including ChatGPT, open-weights models like Mistral, and foundational projects like PyTorch. The team is now focused on pushing the technical boundaries of AI to deliver real value to as many people as possible. Their work combines rigorous engineering with creative exploration to build AI systems that solve impactful problems.
Source: Thinking Machines Lab
Thinking Machines Tinker: Simplifying Fine-Tuning for Language Models