Tinker

Quick Answer

Tinker is Thinking Machines Lab’s managed API for fine-tuning open-weight language models. You write the training loop in Python, and Tinker runs it on managed, distributed GPU clusters, using LoRA to fine-tune efficiently instead of retraining a full model from scratch. It moved from private beta (October 2025) to general availability in December 2025.

Who Makes It

Thinking Machines Lab was founded by Mira Murati, OpenAI’s former chief technology officer, and raised a $2 billion seed round at a $10 billion valuation. Tinker is the company’s first official public product.

What Tinker Is Best For

• Fine-tuning open-weight models without managing infrastructure: no need to provision or babysit a GPU cluster yourself
• Developers who want real control: a low-level training API, not a black-box, no-code fine-tuning tool
• Working with large models: supports large Mixture-of-Experts models like Qwen-235B-A22B
• Cost-efficient customization: LoRA-based fine-tuning uses a fraction of the compute of full fine-tuning

How It Works

You write your training loop locally in Python using Tinker’s API. Tinker handles the distributed execution on managed GPU clusters behind the scenes, so you get the control of writing your own training logic without owning or renting the underlying infrastructure directly. Fine-tuning is primarily done through LoRA (Low-Rank Adaptation), which adjusts a small set of additional parameters rather than every parameter in the base model.

Practical Use Cases

• Fine-tuning an open-weight model on a company’s own data for a specialized task
• Customizing a large MoE model without owning multi-GPU hardware
• Research experimentation that needs real control over the training loop, not just a preset fine-tuning wizard
• Reducing dependency on a single closed model provider by building your own specialized variant of an open model

Honest Limitations

• Not for beginners: writing a training loop assumes real ML engineering knowledge
• Still a paid, usage-based service: managed infrastructure has a real cost
• LoRA has tradeoffs: efficient, but not identical to full fine-tuning in every case
• Relatively new company and product: verify current model support and reliability against your specific use case

Alternatives Worth Knowing

• Hugging Face, for open-weight models and self-managed fine-tuning tooling
• GLM 5.2, Kimi K3, and Poolside Laguna S 2.1, open-weight models you could fine-tune with a platform like Tinker
• OpenRouter, for accessing many models without fine-tuning them yourself

For the broader open-model landscape Tinker fits into, see Frontier Open Models Explained.