Thinking Machines releases Inkling: a 975B open-weights MoE model

Thinking Machines unveiled Inkling, an open-weights Mixture-of-Experts (MoE) transformer model with 975B total parameters (41B active), pretrained on 45 trillion tokens across text, images, audio, and video. It supports a 1M-token context window and balances cost with performance through controllable thinking effort. Alongside Inkling, the company previewed Inkling-Small, a lighter 12B active parameter model with similar capabilities. While not the strongest overall model, Inkling’s combination of multimodal reasoning, efficient inference, and full weight availability makes it a versatile base for fine-tuning and customization.

Inkling is available on Thinking Machines’ Tinker platform for customization, including a new Inkling Playground for developer interaction. In a striking demonstration, the model wrote and executed its own fine-tuning job on Tinker—training itself to become a lipogram that avoids the letter 'e'—completing the pipeline in about 27 minutes. This showcases the platform’s potential for rapid, automated model personalization. Thinking Machines positions Inkling as the first in a family of models, with ongoing work to expand customization for more use cases.