Tinker's December 12, 2025 general-availability release removed the waitlist for Thinking Machines' model-customization service and added image inputs for supported vision-language models. Developers could combine image bytes with text in a training example, allowing fine-tuningTaking a trained model and training it a bit more on your own examples so it gets better at one specific job.Full definition → tasks involving pictures, screenshots or diagrams. Release announcement.
The accompanying image-classification example shows what this enables: give the model an image and train it to produce the category name as text. This extends the customization workflow beyond text-only examples, rather than simply adding an image-viewing interface. Vision input and classifier recipe.
Terms in this piece · Glossary
fine-tuning — Taking a trained model and training it a bit more on your own examples so it gets better at one specific job.
