Now loading...
Thinking Machines has announced the release of Inkling, a groundbreaking AI model designed to enhance human capabilities through customizable intelligence. This innovative model utilizes a Mixture-of-Experts transformer architecture, boasting 975 billion total parameters, with 41 billion active parameters that support a context window of up to one million tokens. Inkling has been pretrained on a staggering 45 trillion tokens encompassing diverse data types including text, images, audio, and video, making it a versatile foundation for various applications.
In addition to the flagship model, the company is previewing Inkling-Small, a more lightweight variant featuring 12 billion active parameters that maintains strong performance at reduced costs and improved latency. Both models showcase robust multimodal capabilities, allowing effective reasoning over text, images, and audio while being highly adaptable for different domains.
With a focus on facilitating customization, Inkling is available for fine-tuning on the Tinker platform, enabling users to refine the model for specific tasks and workflows. To demonstrate this capability, the team conducted a unique experiment where Inkling autonomously managed its own fine-tuning process, illustrating its potential to adapt and improve through interaction.
The company emphasizes its commitment to making customization accessible for a wide range of use cases. New features, such as the Inkling Playground, offer developers an interactive space to engage with the model and refine their applications. The intention is to empower users to tailor Inkling effectively to meet their unique requirements.
Thinking Machines is also keen on addressing real-world challenges through AI and is keen to share its performance statistics, placing significant emphasis on safety and ethical considerations. The model has undergone rigorous evaluations, achieving strong safety ratings that reflect an internal specification for secure AI behavior. According to the firm, Inkling is designed to refuse harmful requests effectively, performing above 98% on the StrongREJECT benchmark.
Inkling’s versatility extends to various applications, including coding and multimedia processing, which the company highlights in its released benchmarks. As part of an ongoing effort to refine and perfect its offerings, Thinking Machines will continue to evolve its model family as user feedback is integrated into development.
For those interested in experimenting with Inkling, the model is available on the Tinker platform, where users can take advantage of an introductory discount. Detailed information regarding pricing and usage can be found in the official documentation.
