Former OpenAI CTO Mira Murati has finally revealed the details of her new project — Thinking Machines Lab. The first public model, named Inkling, is already available for testing and fine-tuning. This is not just another language assistant, but a serious step towards multimodal AI capable of reasoning about text, images, and audio simultaneously.
Inkling's architecture is built on the MoE (Mixture of Experts) principle: only part of the network is activated for each request, ensuring high inference speed without sacrificing analytical depth. The total parameter pool is 975 billion, of which 41 billion are active. The context window is an impressive 1 million tokens. The model was trained from scratch on 45 trillion tokens, including text, images, audio, and video.
However, it is worth noting that despite its multimodal capabilities, Inkling currently only generates text responses — code, structured data, and analytical conclusions. But there is an interesting feature: users can manually adjust the balance between speed and quality using a controlled reasoning level. This provides flexibility rarely seen in open models.
Where Inkling Shows Its Strengths
In my opinion, the model demonstrates its most impressive results in agentic tasks. On the MCP Atlas test, which evaluates task execution using the Model Context Protocol, Inkling scored 74.1% — nearly 30 points higher than Nvidia Nemotron 3 Ultra, the main Western competitor among open models. On SWE-Bench Verified, which tests autonomous bug fixing on GitHub, Inkling also confidently leads: 77.6% versus 70.7% for Nemotron.
On the FORTRESS Adversarial test, which assesses correct handling of malicious requests without excessive blocking, the model scored 78%. This makes it the most powerful open model among Western developments to date. However, as expected, Inkling still lags behind Chinese giants. On Terminal Bench 2.1, which tests autonomous coding in a real environment, Inkling scored 63.8%, while Z.ai GLM 5.2 achieved 82.7%, and the recently released Kimi K3 from Moonshot AI reached 88.3%. Kimi K3 also leads in the PhD-level scientific reasoning test — Humanity's Last Exam.
Thinking Machines positions Inkling as a "general-purpose" model that does not sacrifice performance on some tasks for excellence in others. The bet is on customization: the company believes businesses value a flexible model tailored to a specific task more than a universal "Swiss Army knife."
My expert opinion: Inkling is a serious contender, but it is clearly aimed at niche corporate scenarios where customizability matters. While the model currently lags behind Chinese counterparts in general tests, its strength lies in agentic and coding tasks. If Thinking Machines can scale customization and improve results on scientific tests, Inkling could become a real competitor to DeepSeek and GLM. But for now, this is just the first step.