Tinker
Thinking Machines Lab · Coding · Other
Tinker is a training API from Thinking Machines Lab that fine-tunes language models through LoRA, a method that trains a small add-on adapter instead of rewriting every base weight. You write plain Python to define loss functions, training loops, and data pipelines, and Tinker runs the distributed GPU work behind the scenes. It suits researchers and developers who want control over algorithms and data without managing clusters, and it bills by token usage rather than a flat subscription.
About Tinker
What Is Tinker
Tinker is a managed training API aimed at people who study or build on top of open-source models. Think of it as an AI model fine-tuning API that stays out of your way. It gives you low-level primitives such as forward_backward, optim_step, sample, and save_state, so you decide what the training loop does while Tinker schedules the jobs and keeps the machines running. The pitch is simple: keep your algorithms and your data, hand off the scheduling, tuning, and reliability work that usually eats a research budget.
The service launched in October 2025 and opened to everyone in December 2025, no waitlist. It comes out of Thinking Machines Lab, the company Mira Murati founded after leaving OpenAI as its CTO, and it's built around the team's own research showing that LoRA can match full fine-tuning on learning performance when set up well.
The catch is scope. Tinker trains models that already exist in its catalog, so you aren't pretraining a model from scratch or loading your own weights. You also pay per token, which is cheap for small experiments and adds up fast once you scale a run. Your data stays yours. According to Thinking Machines, it's used only to fine-tune your models and never to train theirs.
Getting Started
- Create an account at the Tinker site and generate an API key from your dashboard.
- Pick a base model from the catalog and install the Python client in your environment.
- Prepare your training data as supervised examples or reinforcement learning environments.
- Call the primitives in a loop to compute gradients, update weights, and sample outputs.
- Save checkpoints with
save_stateand download the weights you want to keep.
Product Information
A quick look at Tinker's pricing, supported platforms, and performance.
Best for
The users, tasks, and scenarios where this tool fits best.
Users
- Machine learning researchers
- Graduate students and lab teams
- Startup engineers
Tasks
- Supervised fine-tuning
- Reinforcement learning runs
- Vision-language tuning
- Distillation
Scenarios
- A research group testing a new post-training algorithm on a fixed budget.
- A team adapting a math or coding model to an internal benchmark.
- A developer who needs a custom model for one narrow task and doesn't want to run a cluster.
Key features
Low-level training primitives
Tinker hands you four core functions. forward_backward runs a forward and backward pass that accumulates gradients. optim_step updates the weights. sample generates tokens for evaluation or RL actions, and save_state saves progress. That small set covers most post-training methods, and you assemble them however your experiment needs.
Swappable base models
Switching from a small model to a much larger one is a string change in your Python code, not a rewrite. No new training script. The catalog covers compact dense models up to large mixture-of-experts systems like Kimi K2 Thinking and the Nemotron 3 line, so you can prototype cheap and scale up later.
LoRA with adjustable rank
Tinker fine-tunes with LoRA, which trains a compact adapter rather than the full weight set. The platform adjusts the adapter's rank on the fly, letting multiple training runs share one GPU pool. Smaller bills for the same experiment. Faster iteration when you're testing ideas.
OpenAI-compatible inference
A standard inference interface mirrors the OpenAI API, so you can sample from a model even while it's still training. Any tool or framework that already speaks the OpenAI format plugs in by pointing at a model path, which saves you from writing glue code.
Managed distributed training
You never touch GPU drivers or cluster setup. Tinker orchestrates the distributed training itself, handling scheduling, resource allocation, and recovery when a job fails. Your time goes into data and algorithms. Not into keeping hardware alive.
Tinker Cookbook
The team ships an open-source library of realistic fine-tuning examples, from classic supervised learning to experimental reinforcement learning pipelines. It's the fastest way to see how the primitives fit together before you write your own loop.
Pros and cons
Pros
- Full control over loss functions and training loops without owning the hardware.
- LoRA keeps compute costs down and lets several runs share one GPU pool.
- Per-token pricing means you only pay for what an experiment actually consumes.
- Model switching is a one-line change, so trying a bigger base model is cheap.
- Downloadable checkpoints keep you from being locked into the platform.
Cons
- No free plan, so you can't test the workflow before spending anything.
- You can't load your own weights or pretrain a model from scratch, which rules out many custom architectures.
- Costs scale with tokens, so a long reinforcement learning run can burn through a budget quickly.
- The primitive API assumes you can write Python and understand training loops; it isn't a point-and-click tool.
Frequently asked questions
It's an API for fine-tuning open-source language models. You supply data and code, Tinker runs the training on its GPU clusters and returns a model you can sample from or download.
Related content
Explore related tools, skills, and articles for Tinker.
Tinker Alternatives
Base Code
Base44 (Wix) · Coding · ProductivityBase Code is a shared cloud development environment from Base44. Teammates describe changes to an existing GitHub codebase in chat, preview them live and send them back as pull requests.
Google EmbeddingGemma 2
Google DeepMind · Writing · Image · CodingGoogle EmbeddingGemma 2 is a 740M-parameter open model that turns text, code, images, audio, and video frames into vectors in one 768-dimension space, running on phones and laptops.
GitGlow
GitGlow · Coding · ProductivityGitGlow is a local-first desktop Git client that lets independent developers review their own code and their AI agents' changes in one place before shipping.