Tinker has officially launched without a waitlist, allowing anyone to sign up and use the platform. Key updates include the introduction of the Kimi K2 Thinking model, which boasts a trillion parameters and is designed for complex reasoning and tool use. Users can now fine-tune this model within Tinker. Additionally, Tinker has improved its inference interface to be compatible with the OpenAI API, allowing for seamless integration with other platforms. The new vision capabilities come from two models: Qwen3-VL-30B-A3B-Instruct and Qwen3-VL-235B-A22B-Instruct. These models can process images alongside text, making them versatile for various applications. Users can input images by combining them with text chunks, which can be particularly useful for tasks like fine-tuning image classifiers. Tinker has already demonstrated the potential of its vision models by successfully classifying images from classic datasets such as Caltech 101 and Stanford Cars. The Qwen3-VL-235B-A22B-Instruct model shows competitive performance even with minimal labeled data, outperforming the traditional DINOv2 model in low-data scenarios. The focus on data efficiency is critical, especially given the scarcity of labeled image data in many real-world applications. Tinker's approach frames image classification as text generation, leveraging the language knowledge of the Qwen3-VL model. This positions Tinker as a practical tool for researchers and developers looking to tackle various vision tasks with limited resources.