Thinking Machines Lab today opened Tinker, a managed API for post-training language models, in a free private beta. The company, founded by former OpenAI CTO Mira Murati, says researchers write their own training loops in Python — including forward_backward and sample primitives that cover most common post-training methods — while Tinker handles scheduling, resource allocation and failure recovery across its internal clusters.

The service supports a range of open-weight models, from small ones up to large mixture-of-experts systems such as Qwen-235B-A22B, with switching between them a matter of changing a single string. Tinker relies on LoRA adapters so multiple runs share the same compute pool, which the company says lowers costs. Alongside the API, Thinking Machines released the Tinker Cookbook, an open-source library with modern implementations of post-training methods for common fine-tuning workflows.

Early users include the Princeton Goedel Team, which trained theorem provers; Stanford’s Rotskoff chemistry group, working on chemistry reasoning; Berkeley’s SkyRL group, running async off-policy RL experiments with multi-agent tool use; and Redwood Research, which ran RL on Qwen3-32B on difficult AI control tasks. Tinker is free to start, with usage-based pricing to come in the next few weeks. It follows July’s $12 billion seed round and gives the startup its first commercial product.