Inkling: Our open-weights model

Mira Murati's Thinking Machines Lab has put out its first open-weights model, and it's a big one. Inkling is a Mixture-of-Experts transformer with 975 billion total parameters, though only 41 billion are active at any given time, and it was trained on 45 trillion tokens spanning text, images, audio and video. It's released under an Apache-2.0 license, and the lab is also promising a smaller sibling called Inkling-Small, at 276 billion total parameters with 12 billion active, though that one's still being tested and its weights aren't out yet.

What's notable is how little Thinking Machines actually says about the thing. The model card is thin compared to what other US labs typically publish, and the Training Data Documentation it links to is even thinner. It basically boils down to two vague paragraphs admitting the training data mixes public domain material with content that might carry intellectual property protections, some pulled from the open internet and public repositories, some acquired from third parties. That's about it.

Thinking Machines isn't pretending Inkling is a frontier model, either. The company says outright it's not the strongest model available, open or closed. Instead the pitch is that Inkling makes a good foundation for customization, thanks to its multimodal abilities, efficient reasoning, and its availability on the company's own Tinker fine-tuning platform. In other words, this is a base you build on, not a model you compete with at the top of leaderboards.

Still, there's plenty to like here. An Apache-2.0 license is genuinely permissive, and the model reportedly holds its own against the open-weight models coming out of China lately. That matters for the US open-weights scene, which gets a new serious player alongside NVIDIA's Nemotron and Google's Gemma line. As for real-world results, a quick test asking Inkling to draw an SVG of a pelican riding a bicycle produced something recognizable, if imperfect. When asked to describe its own rendered image afterward, the model confidently misidentified its bird as a stork or seagull, complete with an oddly detailed rundown of clouds, hills and a bright yellow sun.