Thinking Machines Unveils Inkling, Making Its First Real Bet on Open AI Models

According to a report by TechCrunch, Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati, has released Inkling, its first in-house foundation model — and with it, revealed a strategy that looks markedly different from OpenAI, Anthropic, and Google.

Rather than launching another proprietary chatbot, Thinking Machines has chosen to release Inkling as an open-weight model, allowing enterprises and developers to download, fine-tune, and customize it for their own environments.

Inkling is a 975-billion parameter Mixture-of-Experts model with roughly 41 billion active parameters per inference. It was trained on text, images, audio, and video, although its current outputs remain text-based. According to the company, the model emphasizes calibrated reasoning, allowing it to express uncertainty rather than confidently guessing, while giving users control over the amount of reasoning performed for each task.

Perhaps more importantly, Inkling appears designed to serve as the foundation for Tinker, Thinking Machines' customization platform. Rather than monetizing inference like OpenAI or Anthropic, the company is positioning customization and enterprise adaptation as its primary commercial model.

TechCrunch also notes that Thinking Machines isn't claiming state-of-the-art benchmark leadership. Instead, the company is emphasizing efficiency and adaptability. One internal benchmark suggests Inkling achieves comparable coding performance to NVIDIA's Nemotron 3 Ultra while using roughly one-third the inference tokens, although these results have not yet been independently validated.

The release follows earlier Thinking Machines research into continuous "interaction models," suggesting the company is building an AI platform around collaboration and customization rather than a single universal assistant.

TheMarketAI Take

We've followed Thinking Machines since its inception.

Initially, we questioned whether a multi-billion-dollar valuation built almost entirely on talent could survive the inevitable movement of researchers between OpenAI, Meta, and other frontier labs. More recently, we covered the company's work on multimodal interaction architectures.

Inkling is the first real indication of where all that research has been heading.

What's particularly interesting isn't that Thinking Machines built another large language model. It's that the company appears to be rejecting the prevailing assumption that one frontier model should serve everyone.

Instead, it is betting enterprises increasingly want their own AI, trained on their own knowledge, running under their own control.

That thesis is gaining credibility. Even Microsoft CEO Satya Nadella has recently argued that enterprises risk giving away valuable institutional knowledge every time they rely exclusively on proprietary hosted models.

Whether this becomes a winning business model remains to be seen. Open-weight models can create enormous ecosystems, but they also make monetization more difficult. Success will likely depend less on Inkling itself than on whether Tinker becomes the platform enterprises choose to customize and manage their AI systems.

After months of headlines focused on talent departures, Inkling finally gives investors and the industry something more important to evaluate:

the product.