Tinker (2b4fdb18.connectionism.pages.dev)

Tinker is a newly announced flexible API and managed service for fine-tuning language models that gives researchers low-level control over algorithms and data while abstracting away distributed training complexity. It supports a spectrum of open-weight models—including large mixture-of-experts models like Qwen-235B-A22B—and makes swapping model sizes as trivial as changing one string in Python. The platform handles scheduling, resource allocation, and failure recovery on internal clusters so teams can start small or large runs immediately without managing infra. Technically, Tinker uses LoRA adapters to enable sharing a compute pool across concurrent training runs (lowering cost) and exposes primitives such as forward_backward and sample to express most post-training methods. To help users avoid implementation pitfalls, the team is releasing an open-source Tinker Cookbook with modern, ready-to-run post-training recipes built on the API. Early adopters (Princeton, Stanford, Berkeley, Redwood Research) have used Tinker for theorem proving, chemistry reasoning, multi-agent RL experiments, and RL fine-tuning of Qwen3-32B on control tasks. Tinker is launching in private beta (free to start; usage-based pricing coming), with onboarding beginning immediately and a waitlist for researchers and organizations.