Now loading...
Thinking Machines Lab unveiled significant enhancements to its Tinker platform on Friday, eliminating the waitlist and introducing advanced features for AI model development.
The platform, designed for training and customizing large language models, is now openly accessible to all users. Interested individuals can register via the official sign-up page at auth.thinkingmachines.ai/sign-up. Details on available models and associated costs are listed on the Tinker homepage, while practical coding examples are available in the Tinker cookbook repository.
Among the key additions is the Kimi K2 Thinking model, a trillion-parameter powerhouse optimized for extended reasoning sequences and integration with external tools. Developers can now fine-tune this model directly on Tinker, expanding options for complex AI applications.
Tinker has also launched an inference interface aligned with the OpenAI API, enabling seamless sampling from models even during ongoing training processes. This compatibility allows integration with various OpenAI-compatible ecosystems, streamlining workflows for users. Further technical guidance is provided in the platform’s documentation.
Enhancing multimodal capabilities, Tinker now supports vision inputs through two new models: Qwen3-VL-30B-A3B-Instruct and Qwen3-VL-235B-A22B-Instruct. These tools facilitate the analysis of images, screenshots, and diagrams for diverse uses, including supervised fine-tuning and reinforcement learning. Images are incorporated by combining image data in byte format with text elements, as illustrated in sample code snippets.
To highlight these vision features, the lab shared a cookbook recipe demonstrating how to adapt vision-language models for image classification tasks. Using the larger Qwen3-VL-235B-A22B-Instruct model, which is hosted on Hugging Face, the team achieved solid accuracy with minimal examples per category, with results improving alongside additional training data.
In a practical evaluation, the fine-tuned Qwen3-VL model was tested on four standard image datasets, treating classification as a text generation problem where the AI describes the image’s category. This approach was benchmarked against a conventional vision model, DINOv2-base, augmented with a classification layer. Both underwent low-rank adaptation fine-tuning.
Particularly in scenarios with limited labeled data, starting from just one sample per class, the Qwen3-VL variant surpassed the DINOv2 baseline. Its inherent understanding of language and visuals provides an edge for broader applications beyond simple categorization.
The updates aim to empower researchers and developers in creating cutting-edge AI solutions. The Thinking Machines Lab expressed enthusiasm for upcoming innovations from the community and extended holiday greetings.
