Mira Murati's New AI

A massive 45 trillion tokens of text, image, audio, and video were used to train Inkling, the first in-house AI model from Mira Murati's startup, Thinking Machines Lab. This industry-shaking release represents the startup's first public proofpoint after a year and a half of building its AI infrastructure out of public view.

Unlike flagship closed-source models, Inkling is open-weight, meaning outside developers can download and modify the system directly. The architecture is a mixture of expert system boasting 975 billion total parameters, though it only draws on 41 billion active parameters for any given task to keep runs fast and cheap.

While it reasons natively across multiple formats, its current outputs are restricted to text, code, and structured data. Thinking Machines openly admits Inkling is not the strongest overall model available today, open or closed, but it is highly efficient.

On one coding benchmark, it uses just a third as many tokens as NVIDIA's Neutron 3 Ultra to hit the exact same performance. This launch serves as a high-conviction bet that customizable, open models that organizations can adapt themselves will ultimately outperform the biggest labs' proprietary, one-size-fits-all alternatives.