A service for adapting models
Thinking Machines Lab, Inc. identifies itself as an AI research and product company. Mira Murati announced the company in February 2025. Its work includes both models and a managed service for adapting them, rather than only financing a future model release. [1][2][3]
Tinker launched in October 2025 as a private-beta API for fine-tuning language models. Users controlled training data and algorithms while the company handled distributed training infrastructure. A December 12 update explicitly removed the waitlist. These are separate access stages for a shipped product that preceded Inkling. [3][4]
Research previews and released weights
In May 2026, Interaction models previewed systems that continuously process audio, video and text and respond in real time. The company described a model trained from scratch with a micro-turn design. It was a research preview, rather than unrestricted availability of all demonstrated capabilities. [5]
Inkling followed on July 15 with downloadable weights and hosted fine-tuning through Tinker. The model accepts text, image and audio inputs and generates text. Its card identifies Apache 2.0 licensing; the maximum checkpoint context and the hosted service's supported context settings remain separate specifications. [6][7]
The smaller Inkling-Small was previewed at that launch. The announcement now includes an explicit update that its release followed two weeks later. The larger model's availability should not be projected backward onto a companion still undergoing testing, nor should the July model release erase Tinker's earlier product history. [6]