Thinking Machines Lab releases Inkling, a new multimodal open-weights model

Thinking Machines Lab has released Inkling, a new multimodal open-weights model. It is a Mixture-of-Experts transformer with 975 billion total parameters and 41 billion active parameters. The model is trained on 45 trillion tokens of text, images, audio, and video, and it is released under an Apache-2.0 license. A smaller version, Inkling-Small, is currently undergoing testing and will be released at a later date.

The lab describes Inkling as a base model designed for fine-tuning rather than a frontier model intended to outperform the strongest existing AI. It is optimized for use with the company’s Tinker training platform. While the documentation regarding the specific training data is limited, the lab notes that the model was developed using a combination of public domain content, publicly available internet data, and third-party datasets.

Inkling is capable of processing and generating multimodal content, including the creation of SVG images and the analysis of visual data. By providing a new, viable open-weights option, this release strengthens the US-based open-weights ecosystem, positioning the model as a competitor to existing alternatives like NVIDIA Nemotron and Gemma. Its combination of multimodal functionality and efficient architecture makes it a practical tool for developers looking to customize AI models for specific tasks.