Founded by former OpenAI CTO Mira Murati, Thinking Machines Lab has taken its first major step toward building its own AI systems. The team introduced Inkling, an open-weight model that allows developers not only to use ready-made artificial intelligence but also to adapt it for specific tasks and workflows.
Inkling is built on a Mixture-of-Experts Transformer architecture and includes 975 billion parameters, with 41 billion active during operation. The model can process text, images, audio, and video without relying on separate supporting systems, while its context window reaches up to one million tokens. Alongside the flagship version, the company also introduced Inkling-Small, a more compact model with 12 billion active parameters designed for scenarios where speed and resource efficiency are important.
One of Inkling’s key features is the ability to control how much computing effort the model uses when solving a task. This allows users to receive faster responses for simpler requests while allocating more resources for deeper analysis when needed. The model was also trained for software development, tool usage, multimodal analysis, and following complex instructions.
Thinking Machines has made Inkling available through the Tinker platform for fine-tuning, while the model’s full weights are accessible to developers who want to build their own solutions. The company has also launched a testing environment for the model and says Inkling will become the first release in a new family of AI systems.