Thinking Machines Lab, led by former OpenAI technology chief Mira Murati, unveiled its first artificial intelligence model on Wednesday, marking a notable entry into the growing field of open-weights AI systems. The model, named Inkling, features 975 billion parameters—significantly smaller than the largest proprietary models developed by companies such as OpenAI and Anthropic—but is designed with an emphasis on flexibility and cost efficiency rather than outright performance dominance.

Unlike many advanced closed-source AI models, Inkling’s open weights allow users to modify and customize the model with their own data, supporting a broader movement within the AI industry toward decentralized and more transparent technologies. Thinking Machines framed the release as part of a response to what some industry leaders have criticized as the restrictive “walled garden” strategies employed by frontier AI labs. Executives including Palantir CEO Alex Karp and Microsoft’s Satya Nadella have voiced concerns that dependence on centralized, general-purpose models poses risks to companies’ control over their data and proprietary information.

Inkling was pretrained from scratch using 45 million tokens encompassing text, images, audio, and video, with the training conducted entirely on Nvidia’s advanced computing hardware. The launch follows a multiyear partnership between Thinking Machines and Nvidia, which involved a significant hardware investment aimed at enabling the training and deployment of large-scale AI models. The model’s architecture includes 41 billion “active” parameters—a subset engaged for any given query—allowing for faster and more cost-effective operation.

To support practical customization, Thinking Machines offers Tinker, a cloud-based fine-tuning tool introduced last year, enabling developers and researchers to tailor large AI models without the need for supercomputing resources. Earlier this year, hedge fund Bridgewater Associates utilized Tinker to refine a Chinese open-weights model, Qwen3-235B, on its proprietary data. Bridgewater reported that this fine-tuned version outperformed OpenAI’s GPT-5 and Anthropic’s Claude Opus when handling financial document triage while significantly reducing computing costs.

The company has also emphasized safety, conducting internal tests to assess Inkling’s potential misuse risks—including scenarios involving cyberattacks and biological weapons development—and believes the model performs well under these evaluations. However, given the model’s openness, Thinking Machines acknowledges ongoing efforts to calibrate safeguards, a challenge noted by proprietary AI developers wary of potential vulnerabilities inherent in open-weight models.

On Friday, the company released a manifesto articulating its vision for an AI future grounded in decentralization and the leveraging of localized knowledge. Drawing on the economic theories of Friedrich Hayek, Thinking Machines characterized the prevailing closed-source AI paradigm as analogous to “central planning,” which it argued is insufficient for the nuanced, tacit knowledge that humans apply in everyday tasks. According to the manifesto, centralized aggregation of knowledge is fundamentally limited by the dispersed and often private nature of productive expertise.

As open-weights AI models like Inkling gain prominence, they represent an alternative approach amid intensifying competition between U.S. startups and counterparts in China, such as Alibaba and emerging firms like Z.ai, with many American companies increasingly exploring Chinese open-source offerings for more cost-efficient AI applications. Thinking Machines positions its model as a balance of accessibility, customization, and performance, reflecting a broader shift in the AI industry toward democratizing advanced technologies.