Thinkers at Thinking Machines Lab have unveiled their first homegrown AI model, Inkling, which they’ve open-sourced. Unlike other major players like OpenAI or Anthropic, Inkling allows developers to fine-tune the model for specific tasks. This approach challenges the current paradigm of one-size-fits-all AI solutions.

According to the company’s claims, Inkling is a mixture-of-experts system with 975 billion parameters, though it uses only a fraction of them. It's trained on an enormous amount of data—45 trillion tokens—and can reason across text, images, audio and video. However, its outputs are currently limited to text, such as code, styled artifacts, and structured data.

This model represents the company’s first public proof point after a year and a half of secretive AI infrastructure development. Thinking Machines is marketing Inkling less as a finished product than as a building block for organizations to customize through their platform Tinker. This means that users must ensure the safety and ethical use of the model, rather than leaving it to the company.

The decision to open-source Inkling is part of a broader argument against proprietary AI models. Critics argue that centralizing expertise within one company leads to underperforming AI tools compared to those organizations can adapt themselves. Microsoft CEO Satya Nadella echoed these concerns, warning that enterprises could pay twice as much for subscription fees and business knowledge.

Thinking Machines claims they’re not competing with the best overall models but are focusing on providing a well-rounded performance. This approach has already shown promising results in collaboration with Bridgewater Associates, where an open-source model trained on financial expertise outperformed proprietary AI models while costing less to run.