Inkling: Our Open-Weights Model (thinkingmachines.ai)

Inkling, a newly released open-weights AI model developed from scratch, aims to enhance customization and interactive collaboration within the AI/ML community. This Mixture-of-Experts transformer boasts an impressive 975 billion parameters, with 41 billion activated, and supports a context window of up to 1 million tokens. It was pretrained on a diverse dataset of 45 trillion tokens, including text, images, audio, and video. Inkling emphasizes multimodal capabilities and efficient thinking, making it versatile for various applications across coding, reasoning, and instruction-follower tasks. Alongside it, Inkling-Small—a lighter variant with 12 billion active parameters—demonstrates strong performance with reduced latency and cost. The significance of Inkling lies in its balance of performance, customization, and safety. It is designed to adapt across different workflows, reducing sensitivity to specific tools through randomized training methods. Inkling incorporates features like controllable thinking effort, allowing users to optimize between performance and token efficiency, which is crucial for practical applications. Additionally, it has been extensively tested for trustworthiness and safety, showing strong calibration, instruction-following capabilities, and resilience against censorship. By providing the Inkling Playground via its Tinker console, developers can interactively customize and fine-tune the model, further enhancing its real-world applicability across diverse use cases.