Thinking Machines Lab, a startup founded by several high-profile former OpenAI researchers, has released its first product: Tinker, a service designed to automate the creation of custom frontier AI models. The company says the tool will lower the technical barriers that typically surround large-scale model tuning and put more control in the hands of researchers, engineers, and smaller teams.
“We believe [Tinker] will help empower researchers and developers to experiment with models and will make frontier capabilities much more accessible to all people,” said Mira Murati, cofounder and CEO of Thinking Machines, ahead of the announcement.
Large tech firms and university labs already perform fine-tuning on open source models to tailor them for narrow tasks, from solving math problems to drafting contracts or answering clinical questions. That work normally requires buying or renting racks of GPUs, keeping distributed training stable, and juggling a stack of software for data handling, optimization, and monitoring. Tinker aims to automate many of those steps so a wider group of users can adapt powerful models without managing complex infrastructure.
The tool currently supports fine-tuning of two open source architectures: Meta’s Llama and Alibaba’s Qwen. Developers can call the Tinker API with a few lines of code to start supervised fine-tuning using labeled examples or to apply reinforcement learning, a method that shapes model behavior by rewarding or penalizing outputs. After training, users can export the resulting model and deploy it on their own systems.
The startup’s backers and roster of founders have drawn industry attention. Murati served as OpenAI’s chief technology officer and briefly stepped into the CEO role after the board removed Sam Altman in late 2023; she left the organization about ten months later. She cofounded Thinking Machines with several former OpenAI researchers, including John Schulman, Barret Zoph, Lilian Weng, Andrew Tulloch, and Luke Metz. In July the company disclosed a $2 billion seed round that valued the business at roughly $12 billion.
Schulman led work on applying reinforcement learning to the model that powers ChatGPT, a process that used human feedback to make the system better at dialogue, staying on topic, and avoiding unwanted behavior. He says Tinker gives teams a way to run those training recipes without wiring together their own distributed systems. “There's a bunch of secret magic, but we give people full control over the training loop,” John Schulman said. “We abstract away the distributed training details, but we still give people full control over the data and the algorithms.”
Thinking Machines will accept applications for Tinker access beginning Wednesday. The company is not charging for the API at launch, though Murati and other executives expect to begin monetizing the service at a later date.
A set of beta testers has already been using the platform. Eric Gan, a researcher at Redwood Research, says he has used Tinker’s reinforcement learning features to push models toward producing code backdoors for testing purposes. He argues that training via Tinker can surface capabilities that are hidden behind typical hosted APIs and that adjusting the fine-tuning process is comparatively straightforward. “Tinker is definitely much simpler than doing the RL from scratch,” Eric Gan said, adding: “RL is especially good for if you have a very specialized task and existing models aren't capable of doing it.”
Robert Nishihara, chief executive of Anyscale, which builds tools for operating large AI workloads, praised Tinker’s mix of high-level abstraction and deep configurability. He pointed out that other fine-tuning systems, such as VERL and SkyRL, already exist, but said Tinker’s design will attract wide use. “I think it’s a great API and a lot of people will want to use it,” Robert Nishihara said.
The openness of the models that Tinker supports raises safety questions. One worry among researchers and policymakers is that downloadable open source models can be altered in harmful ways. Thinking Machines currently reviews applicants before granting API access, and Schulman says the company plans to add automated safeguards that will detect and limit misuse over time.
Beyond the product itself, the company has been publishing technical papers on training methods, including techniques to preserve neural network performance during adaptation and to fine-tune large language models with greater efficiency. Those research efforts feed into the Tinker platform, the company says, helping it provide sensible defaults and advanced knobs for people who want to push models in new directions.
The launch also highlights a divide in how AI capabilities are shared. Many U.S. companies hold their most capable models behind hosted APIs, restricting direct access to weights and training routines. By contrast, a growing number of teams in China have released open source models that others can download and modify. Thinking Machines positions Tinker as a way to make model tuning more transparent and available to outside researchers rather than concentrating capabilities inside a handful of firms.
“If you consider what's being done in frontier labs and what other smart people in the world of academia, they're sort of diverging more and more,” Mira Murati said. “And that's not great if you think about how these powerful systems are coming into the world.”
Tinker’s immediate impact will depend on how broadly the startup can scale access while managing the risks of misuse. Its early users point to the platform’s potential to expose latent abilities in models, and the founding team’s pedigree gives weight to the product’s technical claims. The company’s move to offer a user-friendly tuning interface arrives as interest in customizing large models is rising across industry and research groups, and it will likely shape debates over who should be able to change the behavior of the most capable AI systems.