Thinking Machines releases first open AI model, Inkling

AI startup Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, released its first proprietary AI model Wednesday, named Inkling. Unlike flagship models from OpenAI, Anthropic, or Google, Inkling is an open-weight model, allowing external developers and companies to download and modify it directly.

Inkling is a mixture-of-experts system boasting 975 billion total parameters, though it only utilizes a fraction, approximately 41 billion, for any given task. This design approach helps keep the large models faster and more cost-effective to operate. The model was trained on 45 trillion tokens encompassing text, image, audio, and video, and processes these modalities natively, according to company materials.

This release marks the company's first public demonstration of its work over the past 18 months, during which it built AI infrastructure largely out of public view. Earlier, in May, a research preview of "interaction models" was shared – AI systems designed for conversational back-and-forth, including interruptions, unlike typical chatbots. This launch also tests Thinking Machines' core hypothesis: that adaptable AI will outperform generic, one-size-fits-all models offered by major AI labs.

The Inkling model is designed to provide calibrated responses, flagging uncertainty rather than guessing, and allows users to adjust the "thinking effort" to balance speed and quality. The company states that a benchmark shows Inkling using one-third the tokens of Nvidia's Nemotron 3 Ultra to achieve equivalent coding performance. Importantly, Thinking Machines does not claim Inkling is the top-performing model available, explicitly stating in its briefing materials that Inkling is "not the strongest model available today, closed or open."