Tinker

The lab's first product: a managed fine-tuning API that exposes low-level training primitives (forward_backward, optim_step, sample) over open-weight models, so researchers write their own training loops — SFT, RL, whatever — while Thinking Machines runs the distributed infrastructure. LoRA-based, with portable weights users can download. The open-source tinker-cookbook (3.7K+ stars) ships reference post-training recipes. Inkling launched fine-tunable on Tinker day one — the product half of the lab's open-weights + customization thesis.