Control every aspect of model training and fine-tuning
Tinker is a flexible API for efficiently fine-tuning open source models with LoRA. It's designed for researchers and developers who want flexibility and full control of their data and algorithms without worrying about infrastructure management.
Tinker supports 30+ open-source models including Qwen, Llama, DeepSeek, and others ranging from 1B to 397B parameters, allowing users to select optimal model sizes for their research without switching platforms.
Tinker provides a training API that abstracts away infrastructure complexity, allowing researchers to focus on datasets and algorithms. Users control training via four core functions (forward_backward, optim_step, sample, save_state) while Tinker handles distributed GPU orchestration and resource management.
Tinker handles scheduling, tuning, and resource management automatically. Users can save checkpoints and resume training via save_state, enabling rapid experimentation cycles without infrastructure delays.
Tinker uses LoRA (Low-Rank Adaptation) which trains only a small adapter instead of all model weights, matching full fine-tuning performance with significantly lower compute requirements and cost per token.
Tinker's sample function enables token generation for RL actions and evaluation. The API supports models up to 397B parameters, providing sufficient scale for complex RL experiments while maintaining infrastructure abstraction.