Mira Murati’s Thinking Machines Lab Releases Inkling AI Model
By Suad Seferi ·
Mira Murati’s Thinking Machines Lab has released Inkling, the company’s first open-weights artificial intelligence model and its most significant product announcement since the former OpenAI executive launched the startup. Released on July 15, Inkling is a multimodal model capable of working with text, images and audio. It has 975 billion total parameters, with 41 billion activated for each task, and supports context windows of up to one million tokens. Thinking Machines Lab says the model was trained on 45 trillion tokens across text, images, audio and video. Mira Murati Murati, who serves as co-founder and CEO of Thinking Machines Lab, has positioned the company around building AI systems that people and organisations can adapt to their own knowledge, needs and working methods. Inkling follows that direction. Rather than presenting the model as the most powerful general-purpose AI system available, Thinking Machines Lab describes it as a flexible foundation designed for customisation. Developers can download the model weights through Hugging Face, fine-tune it using the company’s Tinker platform and deploy it through several cloud and AI infrastructure providers. The company has been unusually direct about Inkling’s position in the market. It acknowledges that the model does not outperform every leading open or proprietary system. Its value, according to Thinking Machines Lab, is in giving developers greater control over how the model is trained, modified and used. That approach places Inkling in the growing open-weights AI market, where companies release the internal parameters of their models so developers can host, study and customise them more freely than closed systems offered only through an API. Thinking Machines Lab demonstrated the model’s customisation abilities by asking Inkling to create a training process that would teach itself to answer without using the letter “e.” The model generated training examples, prepared an evaluation process and used Tinker to create a specialised version of itself. The demonstration is playful, but the wider point is more serious: Thinking Machines Lab wants organisations to shape AI models around their own expertise rather than relying entirely on a standard model controlled by a single provider. Inkling also includes adjustable reasoning settings, allowing developers to increase or reduce the amount of computation used for a task. This can help balance performance, response time and operating costs depending on how the model is deployed. According to the company’s published benchmark results, Inkling performs competitively across reasoning, coding, instruction-following, vision, audio and tool-use evaluations. It leads some tests and falls behind other open and proprietary models in others. The model is not designed for ordinary consumer hardware. Its size means that most deployments will require specialised infrastructure or commercial cloud services. Thinking Machines Lab has released an optimised version for Nvidia’s Blackwell systems and is working with providers including Together AI, Fireworks, Modal, Databricks and Baseten. The company has also revealed Inkling-Small, a more efficient model with 276 billion total parameters and 12 billion active parameters. Thinking Machines Lab says the smaller version can approach the performance of the full model on several tasks, although its weights have not yet been publicly released. Inkling marks an important moment for Murati, who became one of the most recognisable executives in the AI industry during her time as OpenAI’s chief technology officer. Since establishing Thinking Machines Lab, she has recruited researchers and engineers from several major AI companies and raised substantial expectations about what the new laboratory would build. Until now, much of its public work had focused on research, infrastructure and Tinker, its model-training platform. Tinker is designed to let researchers control fine-tuning while Thinking Machines Lab handles the underlying distributed computing. Inkling is the first clear look at the type of AI models the company intends to produce. It does not immediately reset the competitive landscape, and its performance will still need to be tested independently. But the release establishes Thinking Machines Lab as more than a highly funded research startup built around Murati’s reputation. The company now has its own model, its own training platform and a clear argument: the future of AI may depend not only on who builds the strongest model, but also on who gives people the greatest ability to shape it. Inkling is now available for developers and researchers to explore through Thinking Machines Lab’s official website and Hugging Face release.