Tinker (Thinking Machines)

What it is

Tinker is a managed API for fine-tuning language models, announced by Thinking Machines Lab on 1 October 2025 ^[1]. Fine-tuning means further training an existing model on specific data so that it performs a task better. The service lets users control the algorithms and the data, while Thinking Machines runs the distributed training on its own clusters ^[1][2]. A cluster is a group of linked computers that work together on one large job.

Origins

Tinker was launched alongside the Tinker Cookbook, an open-source collection of post-training methods ^[1]. Post-training is the stage that follows a model's initial training and shapes its behaviour for particular uses.

Main features

Tinker uses LoRA, which lets many training runs share the same compute pool and so lowers the cost of each run ^[1]. It exposes low-level primitives, including forward_backward and sample, ^[1].

At launch the service supported open-weight models including the mixture-of-experts model Qwen-235B-A22B. Changing the size of the model takes a single string change in the Python code ^[1]. Early users included Princeton's Goedel team, a Stanford chemistry group, Berkeley's SkyRL group and Redwood Research ^[1].

Tinker was launched as a free private beta with a waitlist, and usage-based pricing was to follow ^[1].

An application programming interface (API) is a set of rules that lets one program request services from another. For Tinker, this means a developer writes code that sends training instructions to the service, while the hardware and its management stay with the provider. A managed service is one that the provider runs and maintains, so users do not operate the servers themselves.

Its place in development

By August 2026 Tinker supported a wider set of models, including Inkling, Qwen, Nemotron, DeepSeek-V3.1, Kimi K2.6 and gpt-oss ^[2][3]. Inkling launched on Tinker with 64k and 256k context options ^[2][3].