Skip to main content

Thinking Machines Unveils Inkling, an Open-Source AI Model Built for Customization

Thinking Machines Unveils Inkling, an Open-Source AI Model Built for Customization

After a year and a half of quiet development, Thinking Machines Lab—the AI startup founded by former OpenAI CTO Mira Murati—has finally pulled back the curtain. On July 15, 2026, the company released its first self-built open-source AI model, Inkling. And it's not trying to be everything to everyone.

Inkling takes a different approach from the massive, all-purpose models that dominate the headlines. Instead, it's built for businesses that want to tailor AI to their specific needs—without handing over their data to a big tech company.

A Model That Thinks on Demand

Under the hood, Inkling uses a mixture-of-experts (MoE) architecture. That's a fancy way of saying it has a huge brain—975 billion parameters in total—but only turns on the parts it needs for each task, activating about 41 billion parameters at a time. This keeps it both powerful and efficient.

The model was trained from scratch on 45 trillion tokens spanning text, images, audio, and video. Right now, it can output text, code, and structured data. But what really sets it apart is flexibility: users can dial the "thinking intensity" up or down depending on whether they need speed or accuracy.

In code benchmarks, Inkling matched the performance of NVIDIA's Nemotron3Ultra while using only a third of the tokens. That kind of efficiency could translate into real cost savings for companies running AI at scale.

Customization Over Commodity

Thinking Machines isn't aiming to build the next ChatGPT. Instead, Inkling is the foundation for enterprise-grade fine-tuning. The company also launched Tinker, a platform that lets businesses customize the model for their own use cases.

How does Thinking Machines make money? Through revenue-sharing on fine-tuning jobs and managed services. It's a bet that companies will pay for tailored AI solutions rather than settling for a generic model that might not fit their needs.

This strategy aligns with a growing trend: businesses are wary of proprietary data leaks and want AI that can be adapted without losing control. By offering open weights and a customization ecosystem, Thinking Machines is positioning itself as the anti-Google, anti-OpenAI—a company that gives enterprises the keys, not just the car.

What This Means for Open-Source AI

Inkling's launch is a significant moment for open-source AI. It shows that there's room for models that prioritize customization over raw scale. And with a team led by a former OpenAI executive, it brings credibility to the idea that open-source can compete with closed-source giants.

For now, Inkling is available for download and fine-tuning. The real test will be whether businesses embrace it—and whether Thinking Machines can build a sustainable business around customization.

Key Points

  • Inkling is an open-source AI model with 975 billion total parameters, activating 41 billion per task.
  • Uses a mixture-of-experts (MoE) architecture for efficiency.
  • Pre-trained on 45 trillion tokens of text, images, audio, and video.
  • Allows users to adjust "thinking intensity" to balance speed and accuracy.
  • Paired with Tinker, a customization platform monetized through revenue-sharing and managed services.
  • Targets enterprise customization, challenging one-size-fits-all AI models.