Former OpenAI CTO's Thinking Machines launches 'Tinker' to simplify AI fine‑tuning
Mira Murati’s Thinking Machines Lab has introduced Tinker, a managed training API that abstracts distributed infrastructure while giving developers low‑level control to fine‑tune open‑weight models such as Llama and Qwen.
Thinking Machines Lab has unveiled Tinker, a developer tool designed to strip out infrastructure complexity from fine‑tuning large language models while keeping researchers in control of their training loops.
The company announced the product as a managed API for post‑training across open‑weight models.
Low‑level control, managed infrastructure
In its launch materials, the startup has described Tinker as a flexible training API that exposes a small set of primitives—such as forward_backward, optim_step, sample, and save_state—so developers can write their own loops for supervised and reinforcement learning while Tinker schedules and recovers jobs across distributed GPU clusters.
The service has supported fine‑tuning via LoRA adapters, allowing users to download weights for use outside the platform. It has been pitched specifically at researchers and engineers who want fine‑grained algorithmic control without operating GPU fleets.
Model coverage and pricing
Tinker has launched in a private beta with coverage for families such as Meta’s Llama and Alibaba’s Qwen—including large mixture‑of‑experts variants like Qwen3‑235B‑A22B—accessible by switching a model string in code. The company has said Tinker is free to start, with usage‑based pricing to follow.
Instead of prescribing a full training pipeline, Tinker has provided low‑level building blocks so teams can experiment with data, losses and RL strategies while the service abstracts distributed systems, scheduling and failure recovery.
This design has been positioned as a middle path between hands‑off “autotrain” tools and do‑it‑yourself cluster management.
Mira Murati, founder and chief executive of Thinking Machines, has said the goal has been to make frontier‑class model work more accessible to researchers and developers by reducing the operational burden of large‑scale training.
Tinker access has been offered via a waitlist for researchers, universities and organisations; interested users have been invited to request wider access through the company.