Thinking Machines Lab, the AI startup led by former OpenAI CTO Mira Murati, released its first proprietary AI model Wednesday. Called Inkling, it's an open-weight model with 975 billion total parameters that uses a mixture-of-experts design, activating only about 41 billion parameters per task. That keeps it fast and cheap to run. It was trained on 45 trillion tokens spanning text, image, audio, and video, though it currently only generates text, including code and structured data.

The company's bet is that organizations will prefer models they can adapt themselves over the locked-down offerings from OpenAI, Anthropic, and Google. Inkling is designed to give calibrated answers, flagging uncertainty rather than guessing, and lets users adjust how much "thinking effort" the model uses. On coding benchmarks, Thinking Machines claims Inkling uses a third as many tokens as Nvidia's Nemotron 3 Ultra to match its performance. The company explicitly says Inkling is not the strongest model available, but emphasizes well-rounded performance and customizability.

Enterprises can fine-tune Inkling through Tinker, the company's model-customization platform. But customization requires serious machine learning talent, and customers must handle their own safety checks. The approach has backers: Microsoft CEO Satya Nadella recently warned that proprietary models force enterprises to pay subscription costs while also handing over business knowledge embedded in prompts. Hugging Face CEO Clem Delangue predicted most production AI work would shift to private or open-source alternatives. A project with hedge fund Bridgewater Associates found that fine-tuning an open-source model on financial expertise beat top proprietary models on reasoning tests while costing one-fourteenth as much to run. Thinking Machines says it got to market in about nine months, far faster than OpenAI or Anthropic. The company now employs roughly 200 people.

Source: TechCrunch