Chat with Inkling online

Inkling is Thinking Machines Lab's first public model, released on 15 July 2026, and it arrived as an open-weights release rather than a closed API — which is a statement of intent from a lab founded by people who built a lot of what the field runs on.

The architecture is a mixture-of-experts transformer with about 975 billion total parameters and 41 billion active per token, and a context window of up to a million tokens. It is natively multimodal rather than multimodal by adapter: it was pretrained on roughly 45 trillion tokens spanning text, images, audio and video, and images and video frames enter the same decoder as text rather than being translated into it. There are some genuinely unusual design choices inside, including short convolutions in every decoder block and a learned relative-position bias instead of the rotary embeddings almost everything else uses.

What that adds up to in practice is a generalist with unusual reach. A million tokens is enough for a book, a long research archive or an entire codebase in one conversation, and because the multimodality is native you can put a diagram or a recording into the same thread and ask about it directly. The lab also exposed controllable thinking effort, so the model can be pushed to reason harder on the turns that deserve it.

Thinking Machines released a lighter Inkling-Small alongside it. On NinjaChat, Inkling is included in every plan — pick it from the model list and give it something large.