Tinker AI: Fine-Tuning Explained

Tinker is a cloud-based fine-tuning tool from Thinking Machines Lab. It lets a developer at a laptop customize and train large open-weights AI models. You do this without managing the supercomputer that runs underneath.

The company behind Tinker is led by Mira Murati, the former chief technology officer of OpenAI. Tinker was Thinking Machines Lab's first product, released in 2025. It now also powers the lab's own model, Inkling.

This guide explains what Tinker is in simple terms. You will learn how it works, what you can fine-tune with it, and how its pricing model works. We also cover a real proof point from a major hedge fund and the best alternatives to compare.

What Is Tinker?

Tinker is a cloud fine-tuning API built by Thinking Machines Lab. It lets developers customize large open-weights AI models from a laptop. The hard part, running training across many GPUs, is handled for you.

Tinker was the lab's first product. The company is led by Mira Murati, who was chief technology officer at OpenAI. Its goal is to make advanced fine-tuning simple for small teams, not just big labs.

Fine-tuning means teaching a base model your own data, tone, and rules. Tinker gives you that power without a data-center team. You write the training logic, and Tinker runs it on the supercomputing hardware in the background.

How Tinker Works

Tinker works by exposing low-level training primitives while hiding the hard infrastructure. You call simple functions, and Tinker spreads the real work across many GPUs. This is the key trade that makes Tinker useful.

In plain terms, you control the recipe and Tinker runs the kitchen. You decide the data, the method, and the number of training steps. Tinker handles the distributed computing, the scaling, and the failures behind the scenes.

This design gives developers real control without real overhead. You are not locked into a rigid, one-click flow. You can shape how the model learns, then let Tinker do the heavy lifting.

• Exposes low-level primitives like forward passes, backward passes, and sampling.
• Runs distributed training across many GPUs so you do not have to.
• Lets you write custom training logic from a laptop, not a data center.
• Handles scaling, hardware, and infrastructure failures in the background.
• Works as an API, so it fits into a normal developer workflow.

What You Can Fine-Tune With Tinker

Tinker fine-tunes large open-weights model families, so you start from a strong, public base. Supported families include Qwen and Llama, two of the most widely used open models. It now also supports Thinking Machines' own model, Inkling.

Open weights matter here because the model files are public. That lets Tinker load a base model, adjust its weights on your data, and hand back a custom version. Closed models like GPT-5 or Claude do not allow this kind of open training.

In practice, you pick a base model, supply your data, and choose a fine-tuning method. A common approach is LoRA, which trains a small add-on instead of the whole model. This keeps the process fast and cheaper than full training.

Proof It Works: Bridgewater and Qwen3

The clearest proof that Tinker works comes from the hedge fund Bridgewater Associates. Bridgewater used Tinker to fine-tune Qwen3-235B, a large Chinese open-weights model, on its own data. The tuned model was built to triage financial documents.

The result beat both GPT-5 and Claude Opus on that financial-document task. It did so while cutting compute costs by more than 13 times. That is a rare case where an open model, tuned on private data, outperformed top closed models.

This case shows the real value of Tinker for regulated and data-heavy teams. A custom open model can beat a general closed one on a narrow, high-value job. It can also cost far less to run once it is tuned.

Tinker Pricing

Tinker uses a usage-based pricing model, so you pay for the compute you consume. That compute covers both training your model and sampling from it. The more you train and generate, the more you pay.

Thinking Machines has historically offered free credits or a free tier to help new users start. Exact prices for Tinker may not be fully public and can change over time. So we do not list dollar figures that we cannot verify here.

Before you commit, check the official Tinker pricing page for current rates and any free credits. Model your own workload first: estimate your training runs and expected sampling volume. That gives you a realistic cost picture for your project.

Tinker Alternatives

The main alternatives to Tinker are Together AI, OpenAI fine-tuning, Hugging Face AutoTrain, and self-hosted LoRA. Each fits a different mix of control, cost, and ease of use. Your best choice depends on your data rules and skills.

Together AI is the closest direct comparison for open-model fine-tuning, and its pricing is public. OpenAI fine-tuning is simpler but only works on closed models. Hugging Face and self-hosting give the most control at the cost of more setup work.

• Together AI — fine-tunes open models and runs inference, with clear public pricing; the closest head-to-head alternative to Tinker.
• OpenAI fine-tuning — simple and managed, but only tunes closed OpenAI models, so you cannot host or inspect the weights.
• Hugging Face AutoTrain — a low-code path into the huge open-model ecosystem, good for teams that want breadth and community tools.
• Self-hosted LoRA — run fine-tuning on your own GPUs for maximum control and privacy, at the cost of managing all the infrastructure yourself.

Should You Use Tinker?

You should use Tinker if you want to fine-tune open-weights models without running your own GPUs. It fits developers who need real control over training but not the hardware burden. The Bridgewater case shows it can pay off on high-value tasks.

Tinker is less ideal if you only need a quick, closed-model tune with no infrastructure choices. In that case, a managed service like OpenAI fine-tuning may be simpler. It is also worth comparing costs against Together AI before you decide.

For regulated teams, Tinker's open-weights approach is the real draw. You can tune a strong model on private data, then host the result where your rules require. If that fits your needs, Tinker is a strong option to test.