Thinking Machines Challenges AI Giants with Open-Weight Inkling Model

Inkling is a mixture-of-experts system featuring 975 billion total parameters, though it utilizes only 41 billion for any single task. Trained on 45 trillion tokens of text, image, audio, and video, the model is designed to provide calibrated answers while allowing users to adjust "thinking effort" to balance speed and accuracy. The company explicitly states that Inkling is not intended to be the strongest model on the market, but rather a well-rounded starting point for enterprise customization through the Tinker platform.

The release underscores a growing industry debate regarding the utility of centralized, proprietary models. While major labs market general-purpose chatbots, Thinking Machines argues that organizations gain more value by owning and refining their own infrastructure. This strategy aligns with recent warnings from Microsoft CEO Satya Nadella, who noted that enterprises using proprietary models risk inadvertently surrendering their business knowledge to model providers. To prove the efficiency of this approach, the company recently collaborated with hedge fund Bridgewater Associates, where a customized open-source model outperformed top proprietary systems on financial reasoning tests at a fraction of the cost.

Despite these performance claims, the company faces questions regarding its long-term economic model and development practices. While it utilized other open-weight models to assist in early post-training data generation, the startup insists future iterations will be fully self-contained. With a team of approximately 200 employees, Thinking Machines is now positioning itself to capture value through its Tinker platform—focusing on fine-tuning and hosting services rather than the metered access fees that define the current AI landscape.