OpenAI, a former female CTO startup, released its first AI model, drawing on Chinese technology and focusing on cost and performance balance. On July 16, according to the Wall Street Journal, OpenAI’s former chief technology officer Mira Murati is betting on a more customizable AI model, hoping to narrow the gap with her former employer OpenAI

OpenAI, a former female CTO startup, released its first AI model, drawing on Chinese technology and focusing on balancing cost and performance. On July 16, according to the Wall Street Journal, Mira Murati, the former chief technology officer of OpenAI, is betting on more customizable AI models, hoping to narrow the technological gap between cutting-edge AI laboratories such as her former employer OpenAI. Thinking Machines Lab, led by Murati, released the company's first AI model on Wednesday and adopted an "open weight" model, which means other developers can modify the model by combining their own data. This model called Inkling has 975 billion parameters, which is much smaller than the estimated parameters of the most advanced closed-source models of competitors such as OpenAI and Anthropic. Thinking Machines said that the architecture of the Inkling basic model draws on China's DeepSeek-V3 model and is optimized using data generated by Dark Side of the Moon company Kimi K2.5 in the post-training stage. "We trained Inkling into a basic model with broad coverage and balanced capabilities. It has strong performance in many fields and is flexible enough to adapt to different tasks. Whether it is an open source model or a closed source model, Inkling is not currently the model with the strongest overall performance." Thinking Machines said. Thinking Machines said that unlike cutting-edge AI laboratories that focus on improving the raw performance of models, Inkling's design focuses on "stretching a balance between cost and performance." Inkling has nearly 1 trillion parameters, but only 41 billion of them are "active" parameters. This means that only a small part of the AI's "brain" is activated when processing any one user query, making it cheaper and faster to use. The model can also be customized with Tinker, Thinking Machines’ first product launched last year. Tinker is a cloud-based fine-tuning tool for AI developers and researchers. Its goal is to allow developers to customize and train large-scale industrial-grade AI models just by sitting in front of their laptops, without having to worry about the underlying supercomputing infrastructure. Thinking Machines said that the company has conducted security tests on Inkling, including assessing whether it may be used to assist in the creation of biological weapons or help hackers launch cyber attacks. The test results show that the model performed well. However, the company also stated that due to I