Thinking Machines Lab has moved its Tinker platform to general availability while introducing significant new capabilities, including vision processing and enhanced interoperability. The updates remove previous access restrictions and expand the platform's utility for developers and researchers working with advanced models.

Platform Updates

1. General Access: The waitlist requirement has been eliminated,making Tinker immediately available to all users through standard registration. The platform now offers transparent pricing and comprehensive documentation.
2. Enhanced Reasoning Capabilities: Tinker now supports fine-tuning of Kimi K2 Thinking,a trillion-parameter model optimized for complex reasoning tasks and extended tool usage scenarios.
3. OpenAI API Compatibility: A new inference interface provides full compatibility with the OpenAI API standard,enabling seamless integration with existing tools and platforms that utilize this common framework.
4. Vision Processing Integration: The platform has integrated two vision-language models(Qwen3-VL-30B and Qwen3-VL-235B) capable of processing images, screenshots, and diagrams. This enables multimodal applications including image classification, with demonstrated effectiveness even in data-limited scenarios.

Technical Implementation

The vision models process interleaved image and text data, supporting both supervised fine-tuning and reinforcement learning approaches. Early testing shows competitive performance on standard classification datasets including Caltech 101 and Oxford Flowers, with particular advantages in low-data environments compared to traditional vision-only architectures.

Availability and Resources

All features are available immediately through the Tinker platform. The company has released updated documentation, code examples, and specialized cookbooks demonstrating practical implementations, including image classification workflows.

These developments position Tinker as an accessible platform for organizations seeking to customize and deploy advanced multimodal and reasoning models, with particular emphasis on interoperability and practical application development.