Mira Murati has launched Inkling, a new AI model from her Thinking Machines Lab, featuring a unique MoE architecture and impressive performance benchmarks.
Mira Murati, the former Chief Technology Officer of OpenAI, has announced the debut of Inkling, an innovative AI model developed by her new venture, Thinking Machines Lab.
"Today, we are introducing Inkling," Murati stated. "Inkling reasons efficiently across text, image, and audio modalities. We are making the full weights available."
Built using a mixture of experts (MoE) architecture, Inkling activates only a part of its network for each request, enhancing speed while maintaining complexity. The model boasts a remarkable 975 billion parameters and a 1 million token context window, underpinned by training on an extensive dataset of 45 trillion tokens across various media types.
Currently, Inkling delivers text-based outputs, including code and structured data, while allowing users to balance speed and output quality via adjustable reasoning settings. In addition to the main model, a smaller variant called Inkling-Small, comprising 276 billion parameters (with 12 billion active), is in testing, although its full weights are yet to be released.
Performance Highlights
Inkling has demonstrated notable performance, especially in agent tasks, scoring 74.1% on the MCP Atlas benchmark, significantly outperforming Nvidia's Nemotron 3 Ultra. It achieved a score of 77.6% on SWE-Bench Verified, compared to Nemotron's 70.7%. Furthermore, on the FORTRESS Adversarial benchmark, Inkling reached 78%, establishing it as the leading open model from a Western entity, although it still trails behind some Chinese counterparts.
On Terminal Bench 2.1, it scored 63.8%, whereas the Z.ai GLM 5.2 reached 82.7%, and Kimi K3 topped the charts with 88.3%, also excelling in other assessments like Humanity's Last Exam.
Murati envisions Inkling as a "universal" model that excels across various tasks without compromising on others, emphasizing customization to meet diverse business needs. Notably, she departed from OpenAI in September 2024 to focus on her own research endeavors.
