Thinking Machines Lab launches Tinker

Thinking Machines Lab, the company founded by former OpenAI chief technology officer Mira Murati, released Tinker, a managed API for fine-tuning open-weight language models. It was the company’s first public product, roughly a year after its founding and a large seed round attracted attention for a start-up with no product yet released.

Tinker gave developers low-level primitives — functions such as forward_backward and sample — that could be composed to implement most common post-training methods, while Thinking Machines’ infrastructure handled the distributed-training work of scheduling, resource allocation and failure recovery. It supported fine-tuning of a range of open-weight models, including large mixture-of-experts systems such as Qwen-235B-A22B, and used LoRA-based fine-tuning to share compute across multiple concurrent training runs, reducing the cost of each individual run. Researcher Andrej Karpathy praised the design as exposing a useful, low-level slice of the post-training process rather than hiding it behind a fully managed workflow.

Tinker launched in private beta, free to use with a waitlist for access, with usage-based pricing to follow. The product positioned Thinking Machines less as a frontier-model competitor to OpenAI or Anthropic and more as infrastructure for the growing population of researchers and smaller labs doing custom post-training on open-weight base models — a market defined by the difficulty of managing distributed training clusters rather than by access to a proprietary frontier model itself.