Thinking Machines Releases Large-Scale Model "Inkling"

Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, has released a large-scale AI model named "Inkling." Inkling is a multimodal model capable of handling both text and images simultaneously, with a parameter count (the number of adjustable variables used in model training) reaching 975 billion. It is provided in open-weight format, making model weights publicly available, allowing anyone to access and utilize the internal structure of the model.

Open-weight models refer to a format in which trained model data is publicly disclosed, enabling companies and researchers to more easily perform additional training (fine-tuning) on their own data. The proliferation of such publicly available models represents an option for reducing dependency on proprietary models that offer only closed APIs, meeting a certain demand within the developer community. Thinking Machines Lab's prominent positioning of Inkling as a foundational model can be seen as a strategy aligned with this trend.

In terms of performance, it ranks first among U.S. open-weight models on the "Artificial Analysis Intelligence Index," a metric that measures the overall capabilities of AI models. However, when compared with major Chinese open-source models, results show it lags behind on certain tasks. In other words, while it leads in domestic comparisons, Chinese competitors maintain an advantage in some domains within global competition.

Usage fees are set at $1.87 per million input tokens. Tokens are the minimum units by which AI models process text, with approximately one English word corresponding to roughly 1-2 tokens. Thinking Machines Lab positions this pricing not as a standalone model aiming for peak performance, but rather as a foundation for enterprises to customize according to their specific needs.

Thinking Machines Lab is a startup founded by Murati following his departure from OpenAI. His tenure as CTO (Chief Technology Officer) at OpenAI has drawn significant industry attention to the company. The release of Inkling is expected to be the company's first major model unveiled to the public.

The significance of this release extends beyond performance comparison of a single model. The situation where U.S. open-weight models lag behind Chinese competitors in performance reflects one aspect of geopolitical competition in AI development. Whether Inkling can close the gap with China's leading models, and how extensively fine-tuning use cases expand in the future, will become indicators for measuring Thinking Machines Lab's influence.

The open-weight model market is currently an area where diverse players are entering and competition is intensifying. Whether Inkling is adopted by developers as a foundational model depends not only on price competitiveness but also on actual performance after fine-tuning and ease of use. As evaluations accumulate from real-world deployments by enterprise users, Inkling's positioning will become increasingly clear.