Mira Murati's Thinking Machines Lab debuts its first product

Fine-tuning a frontier AI model usually starts with an engineering nightmare. You have to secure GPU clusters, set up brittle distributed frameworks, and pray training runs don't crash halfway through.

Now, Thinking Machines Lab—the heavily backed stealth startup led by former OpenAI CTO Mira Murati—just unveiled its debut product: Tinker, an API designed to automate custom frontier model training.

Applications open Wednesday. The tool is currently free to use, though the company plans to charge for it down the road.

Thinking Machines Lab made headlines in July after raising an eye-popping $2B seed round at a $12B valuation. Its roster reads like an OpenAI all-star alumni list, featuring ChatGPT RL pioneer John Schulman, former research VP Barret Zoph, and safety veteran Lilian Weng.

Reinforcement learning without the DevOps headache

Tinker currently lets developers fine-tune two heavyweight open-weight families: Meta's Llama and Alibaba's Qwen. Users can run both standard supervised fine-tuning and reinforcement learning (RL) with just a few lines of code. Once the job finishes, the custom model can be downloaded and hosted anywhere.

Schulman told WIRED that Tinker handles all the messy plumbing of distributed clusters while leaving developers in full control of their data and training algorithms.

Early beta testers are already praising the developer experience. Anyscale CEO Robert Nishihara noted that while tools like VERL and SkyRL exist, Tinker offers a uniquely clean balance of high-level abstraction and low-level control.

Eric Gan, a researcher at AI safety firm Redwood Research, said he used Tinker's RL features to coax models into writing backdoor-laden code for risk evaluation. Gan noted that building RL infrastructure from scratch is notoriously painful, but Tinker made pushing past the model's vanilla capabilities remarkably straightforward.

Because open-weight models can be turned into dangerous dual-use tools, Thinking Machines is manually vetting beta access for now, with automated safety filters planned for the future.

Why it matters

Frontier AI has become increasingly gated behind proprietary APIs. While labs like OpenAI and Anthropic keep their most advanced reasoning secret, open models like Llama and Qwen give users raw weights, but fine-tuning them with RL still requires elite distributed computing teams.

We think Tinker hits a sweet spot for enterprise teams and independent researchers who want bespoke models for law, medicine, or coding without hiring a dedicated infrastructure team. It shifts the competitive edge from who owns the most GPUs to who has the best domain data and reward signals.

Murati's strategy here is clever: instead of spending billions burning compute to build another base model from scratch, Thinking Machines is positioning itself as the picks-and-shovels platform for the open-source post-training era. But whether Tinker can stay ahead depends on keeping its training efficiency superior to cloud giants like AWS and Microsoft, who are rushing to simplify their own fine-tuning stacks.