Inkling

Inkling is an open-weights, Mixture-of-Experts transformer model from Thinking Machines Lab with 975B total parameters (41B active), pretrained on 45 trillion multimodal tokens. It is designed not as the strongest raw performer but as a balanced, highly customizable foundation model with controllable thinking effort, large context window, and native integration with the Tinker fine-tuning platform. The release emphasizes open access, safety, and the ability for developers to own and adapt the model, demonstrated by a self-fine-tuning loop where Inkling rewrote itself. This positions Inkling as a flexible base for enterprises and researchers who need bespoke AI without vendor lock-in, though it currently trails leading closed models on absolute benchmarks.