VentureBeat
Follow
Thinking Machines' first official product is here: meet Tinker, an API for distributed LLM fine-tuning
Thinking Machines, an AI startup founded by former OpenAI CTO Mira Murati, has launched Tinker, its first product designed to simplify large language model (LLM) fine-tuning. Tinker is a Python-based API that offers developers granular control over training pipelines while managing distributed compute infrastructure. This allows researchers to focus on experimental design and custom model development without infrastructure headaches. The tool provides Python-native primitives, supports various open-weight models including Mixture-of-Experts, and integrates with LoRA-based tuning for cost efficiency. Tinker also includes an open-source companion library, the Tinker Cookbook. Researchers from institutions like Princeton and Stanford have already utilized Tinker for tasks such as formal theorem proving and chemical reasoning, achieving significant performance improvements. The AI research community has praised Tinker for its developer-centric approach, which separates algorithmic control from infrastructure management. Andrej Karpathy highlighted Tinker's clever design, allowing users to retain algorithmic control while offloading infrastructure complexity. John Schulman described Tinker as the infrastructure he always wanted, enabling powerful operations without constant manual oversight. Tinker is currently in private beta and will soon introduce a pay-as-you-go pricing model. Thinking Machines, which raised $2 billion, aims to support open and customizable AI development, distinguishing itself with a focus on multimodal AI systems that collaborate with users.
