Mira Murati Launches First AI Model After OpenAI Departure—Fully Open Source

Introduction of Inkling

Mira Murati, former CTO of OpenAI, has launched a new multimodal AI model named Inkling through her startup, Thinking Machines Lab. This model, featuring 975 billion parameters, is fully open-source and available for download on Hugging Face under an Apache 2.0 license.

Performance Highlights

Inkling has demonstrated impressive capabilities, achieving a score of 74.1% on the MCP Atlas benchmark, which evaluates AI agents' effectiveness in completing real-world tasks. This score significantly surpasses Nvidia's Nemotron 3 Ultra by nearly 30 points, establishing Inkling as the leading open-weights model in the Western market. However, it still faces stiff competition from Chinese models like GLM 5.2 and Kimi K2.6, which excel in certain benchmarks.

Background on Murati and Funding

Murati left OpenAI in September 2024 and launched Thinking Machines Lab in February 2025. The company quickly garnered attention, raising $2 billion in July 2025 at a valuation of $12 billion, marking one of the largest seed funding rounds in Silicon Valley at that time. Although there were plans to raise funds at a $50 billion valuation in November 2025, those discussions fell through by January 2026.

About the Model

Inkling employs a mixture-of-experts architecture, activating only a subset of its network for each task to optimize performance. It can handle text, images, and audio, with a context window capable of processing over 1 million tokens, equivalent to around 750,000 words. The model was pretrained on a vast dataset comprising 45 trillion tokens.

"Introducing our first model, Inkling. Trained from scratch with open weights, fine-tunable on Tinker today," Murati announced on X.

This open-source model provides an alternative for developers in the West who may be hesitant to use Chinese models due to legal and ethical concerns. While it is not the most powerful AI model available, it is considered the most capable open-weights model from a Western lab.

Future Prospects

Thinking Machines also teased a smaller version of Inkling, called Inkling-Small, which contains 276 billion parameters and is expected to match the larger model in reasoning benchmarks. The weights for this version will be released following testing.