Thinking Machines launches Inkling as enterprise AI shifts toward custom models

Thinking Machines Lab has released Inkling, its first internally developed AI model and the clearest public evidence yet of how Mira Murati’s company plans to compete.

What happened

Inkling is an open-weight mixture-of-experts model with 975 billion total parameters, activating roughly 41 billion for each task. Thinking Machines says it was trained across text, images, audio and video, although its current outputs are limited to text, code and structured data.

The company is not claiming that Inkling is the strongest model available. Instead, it is positioning the model as a flexible foundation that organisations can download, adapt and fine-tune through its commercial Tinker platform.

Why it matters

The strategy shifts the value proposition away from a single closed chatbot. Enterprises may care more about controlling deployment, using proprietary data and adapting a model to a specialised workflow than about accessing the largest general-purpose system.

That also creates a clearer commercial question for Thinking Machines: the model weights are open, so the business must make money from training, customisation, infrastructure and support.

The bigger picture

AI competition is splitting into two tracks. Frontier labs are still racing for raw capability, while a second market is forming around smaller or customisable models that companies can govern more directly. Inkling places Thinking Machines firmly in the second camp and gives the company a tangible product against which its unusually high expectations can finally be measured.