Announcing Tinker (thinkingmachines.ai)

Today’s launch of Tinker introduces a managed API for fine-tuning open-weight language models that targets researchers and experimenters who want algorithmic control without managing distributed training infrastructure. Tinker runs on the provider’s internal clusters and handles scheduling, resource allocation, and failure recovery, while exposing low-level primitives like forward_backward and sample so users can compose custom post‑training methods (supervised fine-tuning, RL, async off‑policy, multi-agent loops, etc.). It supports switching between small and very large models (including mixture-of-experts models such as Qwen-235B-A22B) by changing a single string in Python, uses LoRA to multiplex compute across runs and lower costs, and ships an open-source Tinker Cookbook with ready-to-run, modern implementations of post-training algorithms. Early adopters (Princeton, Stanford, Berkeley, Redwood Research) have used Tinker for theorem-proving, chemistry reasoning, complex RL with tool use, and RL on large Qwen models, demonstrating the platform’s flexibility for research workflows. Tinker is entering private beta (free to start; usage pricing coming) with a waitlist and org onboarding available. For the AI/ML community this lowers the barrier to iterate on algorithmic ideas at scale, reproducibly test custom training loops on large open models, and accelerate novel post-training research without building bespoke distributed training stacks.