Thinking Machines Tinker

Tinker is a high-performance training API by Thinking Machines Lab designed for researchers and developers to fine-tune large open-weight models without managing GPU clusters. It abstracts away infrastructure complexities like distributed training, scheduling, and resource allocation while providing low-level control through a simple Python SDK. Users write custom training loops on local machines (CPUs), and Tinker executes the actual computation across powerful GPU clusters. It specifically focuses on Low-Rank Adaptation (LoRA) and supports advanced post-training workflows including Supervised Fine-Tuning (SFT), Reinforcement Learning (RL), Direct Preference Optimization (DPO), and Knowledge Distillation.