Inkling marks Thinking Machines' first model trained entirely from scratch, released as a broad open-weights base intended for customization rather than as a single top-of-leaderboard system. Its architecture is a Mixture-of-Experts transformer with 975 billion total parameters, of which 41 billion activate per token, paired with a context window of up to one million tokens. The pretraining corpus spans roughly 45 trillion tokens drawn from text, images, audio, and video, giving the model native multimodal grounding. Full weights are published on Hugging Face, and the model is also offered for fine-tuning through Thinking Machines' Tinker platform, signaling a deliberate emphasis on downstream adaptation by outside teams rather than a closed, turnkey experience.

In practice, Inkling is positioned for users who want to adapt a large foundation model to their own tasks rather than pick a finished product off the shelf. Its native reasoning covers text, images, and audio, and it exposes controllable thinking effort so developers can tune the depth of deliberation against latency and cost. The release also includes a preview of Inkling-Small, a 276-billion-parameter sibling with 12 billion active parameters trained with a similar recipe, giving teams a lighter option when the full-scale model is more capacity than they need. Together, the weights release, fine-tuning pathway, and smaller sibling make Inkling best suited to research groups and product teams comfortable doing their own post-training.