Thinking Machines Launches Open AI Model Inkling to Challenge Closed Enterprise AI

Thinking Machines Lab has introduced its first in-house artificial intelligence model, Inkling, marking the startup’s first major product release since emerging from stealth. Founded by former OpenAI Chief Technology Officer Mira Murati, the company is taking a different approach from many leading AI developers by releasing Inkling as an open-weight model that organisations can download, modify, and fine-tune for their own needs.

The launch reflects Thinking Machines’ broader strategy of challenging proprietary AI platforms with technology that gives enterprises greater control over how models are adapted, deployed, and improved. Rather than positioning Inkling as the industry’s most powerful model, the company says it is focused on delivering balanced performance and customisation for business users.

Open-Weight AI Designed for Enterprise Customisation

Inkling is built using a mixture-of-experts architecture with 975 billion total parameters, although only around 41 billion are active during any given task. This design allows the model to deliver high performance while reducing computing costs and improving efficiency.

According to Thinking Machines, Inkling was trained on approximately 45 trillion multimodal tokens spanning text, images, audio, and video. While the model can reason across all four data types, its current public release generates text-based outputs, including software code, structured data, and formatted documents.

The company has also introduced features intended to improve reliability, including calibrated responses that acknowledge uncertainty instead of generating potentially inaccurate answers. Users can also adjust the model’s "thinking effort” to balance response quality and processing speed based on the workload.

A Different Strategy from OpenAI, Anthropic and Google

Unlike proprietary AI services such as ChatGPT, Claude and Gemini, Inkling is designed as a foundation that enterprises can customise themselves through Thinking Machines’ Tinker platform. Businesses can fine-tune the model using their own internal knowledge and workflows rather than relying entirely on centrally trained AI systems.

The company argues that organisations often possess specialised expertise that general-purpose AI models cannot easily capture. By allowing customers to adapt Inkling for their own industries and operations, Thinking Machines believes businesses can achieve stronger performance than with one-size-fits-all commercial AI assistants.

The approach also shifts responsibility for customised deployments to customers, who must ensure their fine-tuned models remain safe and appropriate for production use.

Growing Industry Interest in OpenAI Models

The release comes as interest in open-weight and open-source AI continues to grow across the technology industry. Microsoft Chief Executive Officer Satya Nadella recently argued that enterprises relying exclusively on proprietary AI systems may effectively pay twice—once through subscription costs and again by contributing valuable business knowledge through prompts and feedback.

Hugging Face Chief Executive Officer Clem Delangue has also predicted that while frontier AI models will continue serving research and high-value applications, many production workloads will increasingly move toward private or open alternatives that organisations can control themselves.

Thinking Machines believes this trend creates an opportunity for platforms such as Inkling, where the underlying model remains openly available. At the same time, value is generated through enterprise training, fine-tuning, hosting, and deployment services.

Early Enterprise Results Highlight Potential

One example highlighted by the company involved a collaboration with Bridgewater Associates, where researchers trained an existing open-source model using the hedge fund’s financial expertise. According to Thinking Machines, the resulting model achieved stronger financial reasoning performance than leading proprietary AI systems while operating at significantly lower cost. The evaluation, however, was conducted internally by the participating organisations rather than through an independent benchmark.

Thinking Machines also reports that Inkling can achieve comparable coding performance using roughly one-third of the tokens required by Nvidia’s Nemotron 3 Ultra open-weight model, demonstrating the company’s emphasis on computational efficiency alongside model quality.

Building the Business Around Tinker

Although Inkling itself is freely available as an open-weight model, Thinking Machines appears to be building its long-term business around Tinker, its enterprise model-customisation platform. The company expects revenue to come from training, fine-tuning, hosting and deployment services rather than charging customers for every model interaction.

The startup has partnered with Nvidia to deploy Vera Rubin computing infrastructure and trained Inkling using Nvidia GB300 NVL72 systems. However, it has not disclosed the total cost of training the model or detailed its long-term funding strategy.

Thinking Machines now employs around 200 people and continues expanding its engineering organisation despite several high-profile departures earlier this year. The company says its internal culture prioritises collaboration and continuity over reliance on individual personalities, even as public attention remains closely tied to founder Mira Murati.

With Inkling, Thinking Machines is positioning itself as a significant new player in enterprise artificial intelligence by betting that flexible, customisable AI models will become increasingly important as organisations seek greater ownership over their data, workflows and AI capabilities.