Thinking Machines Lab Releases Inkling: A 975B-Parameter Open-Weights Multimodal MoE With 41B Active Parameters And Controllable Thinking Effort

Editor's take

Thinking Machines Lab has introduced Inkling, an open-weights multimodal model featuring 975 billion total parameters, with only 41 billion active during inference, and a context window of one million tokens.

This release is significant as it pushes the boundaries of accessible, large-scale multimodal AI, offering a powerful tool for researchers and developers without the prohibitive costs often associated with proprietary models. The Apache 2.0 license ensures broad adoption and further innovation in areas like complex reasoning and long-form content generation, potentially democratizing advanced AI capabilities.

Future developments to monitor include Inkling's performance on benchmarks against models like Google's Gemini 1.5 Pro and OpenAI's GPT-4V, and how effectively its "controllable thinking effort" feature can be leveraged for efficiency gains and specialized applications. The long-term impact will depend on the community's ability to fine-tune and deploy Inkling for diverse multimodal tasks.