Thinking Machines Lab launches Tinker, a fine-tuning API

What happened

On 1 October 2025 Thinking Machines Lab announced Tinker, its first product ^[1]. Tinker is a managed API for fine-tuning open-weight language models ^[1]. Open-weight means the trained numbers inside the model, its weights, are published so that others can download and use them. Fine-tuning means running extra training on a model that already exists, so that it becomes better at a particular kind of task.

How it works

Tinker is an API, a service that users reach through software rather than through a chat window. In a managed service of this kind, the provider keeps the large machines and the software that drives them, and the customer focuses on the data and the goal.

Tinker uses LoRA, a method that trains a small set of added parameters instead of changing every weight in a model. The announcement said LoRA lets the service share compute across training runs ^[1]. In practice, that means several users' jobs can be run on the same hardware more efficiently than if each job rebuilt the full model.

What it supports

Tinker supported large mixture-of-experts models such as Qwen-235B-A22B ^[1]. A mixture-of-experts model is built from many smaller expert subnetworks, and only a subset of them is used for each input. The name Qwen-235B-A22B is the model's own label from its developers, and the announcement named it as an example of what Tinker could train ^[1].

Launch and early users

Tinker launched as a free private beta, which means access was limited to invited users during the trial period. Usage-based pricing was planned for later ^[1]. Early users included groups at Princeton, Stanford and Berkeley, and Redwood Research ^[1].

Why it matters

Tinker was the first product from a company that had been founded only months earlier. It offered researchers and developers a single interface for fine-tuning large open models, without having to build and run the training infrastructure themselves. For readers following the field, it is an example of a model-building lab offering its tools to others as a service, in addition to publishing research.