Thinking Machines Lab, an AI startup led by former OpenAI executive Mira Murati (미라 무라티), has unveiled its first in-house AI model, Inkling.

It is an open-weight model that outside developers and companies can download and modify, a different approach from OpenAI, Anthropic and Google, which focus on closed models.

TechCrunch reported on July 15 that Inkling marks the first time Thinking Machines Lab has publicly revealed AI infrastructure it built in secret for about 18 months.

Inkling is a mixture-of-experts model with 975 billion parameters. It uses about 41 billion parameters in practice for each task. It was trained on 45 trillion tokens across text, images, audio and video. The company said it can reason across text, image, audio and video formats.

Thinking Machines Lab is emphasising an enterprise custom-AI strategy rather than a race for top performance. It is based on a view that AI tuned by each organisation can deliver better results than having all customers use the same model. Inkling is closer to a starting point than a finished service. Companies can do additional training and fine-tuning via the model customisation platform Tinker. They must also take responsibility for safety management, and fine-tuning requires a high level of machine-learning expertise.

Inkling flags uncertainty and lets users adjust “thinking intensity” between speed and the amount of reasoning.

Thinking Machines Lab said Inkling is not currently the strongest model among public and private models. It said that in one benchmark it matched Nvidia Nemotron 3 Ultra in coding performance while using about one-third as many tokens.

Thinking Machines Lab said it pre-trained Inkling in-house from the start, but acknowledged it used other open-weight models when creating some of the initial follow-on training data before large-scale reinforcement learning. That included Moonshot AI’s Kimi K2.5. It said that starting with the next model, it will carry out follow-on training fully within its own system.

Its monetisation plan is not yet clear. TechCrunch reported that Thinking Machines Lab is structured to earn revenue less from the model itself and more from fees in an ecosystem for training, fine-tuning and hosting via Tinker. Thinking Machines Lab has about 200 employees.