Tinker

Tinker - Thinking Machines Lab

Overview

Tinker is a training API tailored for researchers and developers, aimed at simplifying and enhancing the process of model training and fine-tuning. It allows users to focus on their ideas, datasets, and algorithms while managing the complexities of infrastructure. The API is designed to perform efficient fine-tuning of a variety of machine learning models, using innovative techniques like Low-Rank Adaptation (LoRA).

Features

Key Functionalities

Tinker embodies four essential functions that empower developers and researchers:

1. forward_backward: This function performs a forward and backward pass through the model to accumulate gradients.

2. optim_step: It updates the model's weights based on the accumulated gradients from the training process.

3. sample: This allows users to generate tokens for diverse applications, including interaction, evaluation, or reinforcement learning actions.

4. save_state: This functional capability enables users to save the training progress so that they can easily resume later without the need to restart.

Supported Models

Tinker supports a wide array of models, which include:

• QWEN Models:

  • Qwen3-4B-Instruct-2507
  • Qwen3-8B-Base
  • Qwen3-30B-A3B-Base MoE
  • Qwen3-235B-A22B-Instruct-2507 MoE

• LLAMA Models:

  • Llama-3.2-1B
  • Llama-3.1-70B Dense
  • Llama-3.3-70B-Instruct Dense

• GPT-OSS Models:

  • GPT-OSS-120B MoE
  • GPT-OSS-20B MoE

• DeepSeek Models:

  • DeepSeek-V3.1 MoE

This broad selection facilitates flexibility in model selection for various research and application needs.

Efficient Fine-Tuning with LoRA

Tinker utilizes Low-Rank Adaptation (LoRA) for fine-tuning models. This innovative method focuses on training a small adapter added to the model rather than altering all original weights. As a result, it achieves comparable learning performance to full fine-tuning while requiring significantly less computational power and providing greater flexibility.

User Testimonials

Numerous users have praised Tinker for its ability to streamline the research process:

• Tyler Griggs from Berkeley emphasized Tinker's capacity to let researchers concentrate on data and algorithms without worrying about the complexities of infrastructure.

• Ziran Yang and colleagues from Princeton highlighted that Tinker alleviates the engineering overhead, allowing them to focus on their core research activities.

• Jason Liu from Stanford remarked on the abstraction of training infrastructure, which fosters easier engagement in reinforcement learning work.

• Eric Gan from Redwood Research commended Tinker for its reliability and efficiency in quickly iterating through research processes without the burden of hardware management.

FAQs

How to Access Tinker

To get started, individuals can sign up for the waitlist. Organizations and universities interested in broader access are encouraged to contact Tinker directly.

Purpose and Audience

Tinker is specially designed for researchers and developers looking for a flexible API to fine-tune open-source models efficiently. By emphasizing user control over data and algorithms, Tinker liberates its users from the challenges associated with infrastructure management.

Handling of Data

Training data is strictly used for fine-tuning models and is not used to train Tinker's own models. Users also have the ability to download their model weights whenever necessary.

Cost Structure

The pricing model for Tinker is based on compute usage, with specified costs for different models and various functions like training and sampling. Below are some examples of the pricing structure:

Model	Prefill	Sample	Train
Qwen3-4B-Instruct-2507	$0.07	$0.22	$0.22
Llama-3.1-70B	$1.05	$3.16	$3.16
GPT-OSS-120B	$0.18	$0.44	$0.52

Conclusion

Tinker stands as a vital tool in the arsenal of those engaged in AI and machine learning research. By offering a comprehensive, user-friendly API for model training, it allows researchers to dedicate their efforts toward innovation rather than logistical overhead. Whether you are an individual developer or part of a larger research organization, Tinker provides the flexibility, control, and infrastructure support needed to accelerate the exploration of artificial intelligence. The integration of advanced fine-tuning techniques such as LoRA further positions Tinker as a forward-thinking solution for the evolving demands of machine learning development.