Inkling: Our open-weights model

Mira Murati's Thinking Machines Lab has unveiled Inkling, a significant new open-weights artificial intelligence model designed to advance the accessibility and democratization of large language models. The release represents a notable development in the AI landscape, where open-source alternatives to proprietary systems continue gaining traction among researchers and developers worldwide.

Inkling is a Mixture-of-Experts transformer architecture featuring 975 billion total parameters with 41 billion active parameters. The model is licensed under Apache 2.0, enabling broad commercial and research applications. The training dataset encompasses an impressive 45 trillion tokens spanning text, images, and audio modalities, positioning Inkling as a comprehensive multimodal system capable of processing diverse input types. This architecture represents an efficient approach to scale, activating only a subset of parameters during inference while maintaining the model's total capacity.

• Open-source accessibility: The Apache 2.0 license removes licensing barriers, allowing organizations to deploy, modify, and commercialize the model without proprietary restrictions
• Multimodal capabilities: Support for text, images, and audio creates opportunities for more versatile AI applications across different industries
• Efficient scaling: The Mixture-of-Experts approach reduces computational requirements during deployment while maintaining model capacity
• Competitive landscape: The release intensifies competition with proprietary models, potentially accelerating innovation in open-source AI development
• Research democratization: Academics and smaller organizations gain access to state-of-the-art model architecture for experimentation and advancement

The release of Inkling highlights the continued momentum toward democratizing advanced AI capabilities. As larger models become increasingly available through open-source channels, the barrier to entry for organizations developing AI applications diminishes significantly. This shift encourages innovation beyond well-capitalized tech companies and enables researchers to build upon cutting-edge architectures without proprietary constraints. Inkling's multimodal nature and efficient parameter activation address real deployment challenges while maintaining competitive performance, potentially reshaping how organizations approach AI infrastructure decisions moving forward.