Thinking Machines presents its first AI model “Inkling”.

The AI ​​start-up Thinking Machines, founded by former OpenAI technical director Mira Murati and recently valued at twelve billion US dollars, has published its first AI model called Inkling. Unlike the top models from Anthropic, OpenAI and Google, the company openly provides the trained weights, i.e. model parameters. This allows developers and companies to operate Inkling on their own hardware or in a cloud of their own choosing and adapt it to their respective requirements.

Inkling uses a so-called Mixture of Experts architecture (MoE), in which the entire model does not work at the same time. Of a total of 975 billion parameters, only 41 billion from specialized subnets are active for each token processed. This reduces the computing effort and thus also response times and operating costs.

Inkling was pre-trained with 45 trillion tokens made up of text, images, audio and video data. Inkling can process up to a million tokens at the same time within one request. According to Thinking Machines, the model evaluates information from texts, images and audio data together and derives conclusions from it. Inkling currently produces text-based content, including body text, program code, structured data, and code for formatted documents and web applications.

Bet on customizability

Independent tests by the analysis platform Artificial Analysis paint a mixed picture. Inkling is currently the most powerful open-weight model from the USA, but lags behind the best open-weight models from China in the overall ranking.

Thinking Machines admits that Inkling does not achieve the highest overall performance among either open or closed models. Instead, the company positions Inkling as an adaptable basic model that can be tailored to specific tasks using its own fine-tuning platform Tinker, among other things. This is also the strategic alternative to the large AI providers: The start-up assumes that for many companies, in the long term, the most powerful universal model is less important than one that they can control themselves and tailor to specific use cases using their own data. In addition to Tinker, Inkling is also available from cloud and API providers and as a model download on Hugging Face.

Thinking Machines was founded in 2025 by former OpenAI technical director Mira Murati and employs numerous former employees of leading AI laboratories. Even before releasing its own product, the start-up raised two billion US dollars. Inkling’s announcement was preceded by the launch of the fine-tuning platform Tinker at the end of 2025 and the first presentation of an interaction model in May 2026.

(tobe)