Thinking Machines, founded by ex-OpenAI CTO Mira Murati, has launched Tinker, an innovative API designed for training language models. Tinker aims to democratize AI development by enabling researchers and developers to build and customize open-source models without hefty computational resources. The API supports various open-weight models from Meta and Alibaba, including the large-scale Qwen-235B-A22B, allowing users to switch models effortlessly with simple code updates.
Tinker operates as a managed service on proprietary compute clusters, handling resource management, fault tolerance, and scheduling seamlessly. Utilizing LoRA (Low-Rank Adaptation), it facilitates parallel fine-tuning, optimizing costs effectively. Currently in closed beta, Tinker is free to early users, with plans for usage-based pricing soon. The Tinker Cookbook aids developers in avoiding common fine-tuning pitfalls, highlighting the platform’s commitment to enhancing AI model flexibility. This focus on open-weight models indicates a strategic shift towards accessible AI solutions over closed models like GPT-5.
Source link
