Thinking Machines’ Tinker lets developers fine-tune large language models from a laptop, with an open source Cookbook making post-training experiments accessible to academics and hobbyists alike.
Thinking Machines, the AI startup founded by former OpenAI CTO Mira Murati, has launched its first product, Tinker, an API that allows developers to fine-tune large language models (LLMs) on standard hardware.
The managed service handles scheduling, resource allocation, and failure recovery, letting users start small or large training runs without managing infrastructure. Tinker supports all popular open-weight AI models, including Alibaba’s Qwen and Meta’s LLaMA, as well as mixture-of-experts (MoE) models. Developers can write training loops in Python on their laptops, with Tinker executing them on distributed GPUs.
Tinker utilises LoRA, a method that fine-tunes large models efficiently by adding ‘lower-rank’ matrices, enabling adaptation without modifying entire models. Its low-level API primitives, including `forward_backward` and `sample`, facilitate post-training methods.
A major open-source component, the Tinker Cookbook, provides modern implementations of post-training methods on the Tinker API, lowering the barrier for academic and hobbyist experimentation. Research groups from Princeton, Stanford, Berkeley, and Redwood Research have already begun using the platform. Berkeley’s SkyRL group, for example, ran experiments on a custom async off-policy RL training loop with multi-agents and multi-turn tool-use.
John Schulman, co-founder of OpenAI: “Tinker provides an abstraction layer that is the right one for post-training R&D.”
Horace He, Thinking Machines: “Sadly, these factors all push fine-tuning/RL out of reach of hobbyist setups.”
Xi Ye, Princeton University: “RLing >10B models on a typical academic setup… is a hassle, but with Tinker I can focus more on the data/algorithms.”
Tyler Griggs, UC Berkeley: “I don’t know of an alternative product that provides this… The API design is clean… and could easily express each of these in the Tinker API.”
Tinker is currently available on a waitlist and free to start, with usage-based pricing to follow.
