Thinking Machines, the AI startup founded by the former OpenAI CTO Mira Murati, has announced its first product, Tinker.
Tinker is an API that helps developers fine-tune large language models. “It empowers researchers and hackers to experiment with models by giving them control over the algorithms and data while we handle the complexity of distributed training,” Thinking Machines said in a blog post.
Tinker is a managed service that runs on the company’s training infrastructure. The service handles scheduling, resource allocation and failure recovery. “This allows you to get small or large runs started immediately, without worrying about managing infrastructure,” said the company.
The API supports all the popular open weights AI models from Alibaba (Qwen) and Meta (Llama), ranging from small models to large mixture-of-experts (MoE) models. With the API, Thinking Machines says it is now possible to “write training loops in Python on your laptop,” while Tinker will run them on its distributed GPUs.
Tinker utilises LoRA, a method that fine-tunes models efficiently by adding ‘lower–rank’ matrices. This approach enables large models to adapt to specific tasks by attaching lightweight components rather than modifying the entire model.
Tinker’s API provides low-level primitives such as forward_backward and sample, which can be used to implement the most common post-training methods. “Even so, achieving good results requires getting many details right,” said the startup.
The startup has also released an open-source library called the ‘Tinker Cookbook’, which details modern implementations of post-training methods that run on top of the Tinker API.
Thinking machines has said that groups of researchers from Princeton, Stanford, Berkeley and Redwood Research have already been using Tinker. “Berkeley’s SkyRL group ran experiments on a custom async off-policy RL training loop with multi-agents and multi-turn tool-use,” said the startup.
Tinker is currently available on a waitlist and is free to start, with usage-based pricing to be introduced in the coming weeks.
Why Tinker
“Tinker provides an abstraction layer that is the right one for post-training R&D,” said John Schulman, the co-founder of OpenAI, who now works at Thinking Machines.
Meanwhile, Horace He, from Thinking Machines, explained in a post on X that one fundamental reason for Tinker to be released is the rise of MoE models, as they require large multinode deployments.
ALSO READ: OpenAI’s New Partnership with Shopify, Etsy will Allow ChatGPT to Sell Products
He added that GPUs achieve good performance only with large batch sizes (over 256 tokens), but MoE routing increases the parallel request requirement dramatically. For example, with DeepSeekV3’s 32-way sparsity, efficiency needs around 8,192 parallel requests.
“Sadly, these factors all push fine-tuning/RL out of reach of hobbyist setups,” said He, underscoring the need for Tinker.
Several developers and researchers have already had the opportunity to work with Tinker and shared their experiences. The consensus appears to be that this enables a greater focus on algorithms and data for the AI model, while leaving the infrastructure-related tasks to Tinker.
“As an academic, I find it an amazing platform that makes RL training at >10B scale easily accessible. RLing >10B models on a typical academic setup (single node, a few GPUs) is a hassle, but with Tinker I can focus more on the data/algorithms (sic),” said Xi Ye, a postdoctoral fellow at Princeton University, in a post on X.
Tyler Griggs, a PhD student at the University of California, Berkeley, shared his initial impressions in a post on X, echoing a similar sentiment. “I don’t know of an alternative product that provides this,” said Griggs, indicating how it helps developers ‘ignore’ the complexities of compute and infrastructure.
“The API design is clean. I toyed with multi-turn RL, async RL, custom loss functions, even some multi-agent training, and could easily express each of these in the Tinker API,” he added.
ALSO READ: OpenAI to Challenge TikTok and Instagram With Sora 2
