Tinker API Simplifies Model Fine Tuning | Generated by AI
Tinker is a new flexible API launched by Thinking Machines AI for fine-tuning language models. It’s designed to let researchers, developers, and “hackers” experiment with open-weight models (both large ones like Qwen-235B-A22B and smaller ones) while giving them full control over algorithms and data. The service handles the heavy lifting of distributed training—like scheduling, resource allocation, and failure recovery—on their internal clusters, using techniques like LoRA to make it efficient and cost-effective.
Key highlights:
- Ease of use: Switch models with just a string in Python code.
- API primitives: Low-level tools like
forward_backwardandsamplefor common post-training methods. - Open-source support: Comes with the Tinker Cookbook library on GitHub, featuring modern implementations of fine-tuning techniques.
- Purpose: Simplifies research on cutting-edge models, reducing infrastructure headaches so you can focus on innovation.
It’s currently in private beta (free to start, with usage-based pricing coming soon), and you can join the waitlist at thinkingmachines.ai/tinker. Early users include teams from Princeton, Stanford, Berkeley, and Redwood Research.