Tinker: General Availability and Vision Input (thinkingmachines.ai)

Tinker has officially launched its updates, moving from a waitlist system to general availability and introducing significant enhancements. The new Kimi K2 Thinking model, boasting a trillion parameters, allows users to perform complex reasoning and tool use. Additionally, Tinker now features an OpenAI API-compatible inference interface, enabling seamless integration with existing OpenAI platforms for efficient model sampling. Crucially, Tinker supports vision inputs through two new models: Qwen3-VL-30B-A3B-Instruct and Qwen3-VL-235B-A22B-Instruct. These models facilitate the processing of images, screenshots, and diagrams, enhancing capabilities across various applications. Demonstrating this, Tinker has finetuned Qwen3-VL-235B-A22B-Instruct for image classification on well-known datasets, outperforming traditional models like DINOv2 in low-data scenarios. This integration of language and vision abilities positions Tinker as a versatile tool for developers and researchers, opening new pathways for efficient data utilization in AI-driven tasks.