StorySoftware
Thinking Machines launches Tinker, an API for fine-tuning open-weight models
Thinking Machines launched Tinker in private beta, an API that lets researchers fine-tune open-weight models from small to large mixture-of-experts like Qwen-235B while it handles distributed training. It uses LoRA, which trains a small add-on set of weights, so many jobs can share the same GPUs; it was free during the beta.
- LoRA approach enables cost sharing across multiple training runs on the same GPUs; low-level primitives like forward_backward and sample for post-training.
- Early adopters: research groups at Princeton, Stanford, Berkeley, and Redwood for math theorem proving, chemistry reasoning, and RL.
- Open-source Tinker Cookbook provides modern post-training implementations; usage-based pricing after beta phase.