Thinking Machines makes its Tinker AI fine-tuning service generally available – News

Thinking Machines Lab Inc. today launched its Tinker artificial intelligence fine-tuning service into general availability.

San Francisco-based Thinking Machines was founded in February by Mira Murati (pictured), the former chief technology officer of OpenAI Group PBC. During her time at the AI provider, Murati oversaw the development of ChatGPT, Sora and other products. The Thinking Machines team also includes other prominent AI experts. The company recruited PyTorch co-creator Soumith Chintala from Meta Platforms Inc. last month.

Thinking Machines raised a $2 billion seed round in June at a $10 billion valuation. The investment included the participation of Nvidia Corp., Advanced Micro Devices Inc., ServiceNow Inc. and other prominent backers. The company debuted Tinker four months later.

Tinker is a cloud service designed to reduce the amount of effort involved in fine-tuning large language models. Fine-tuning is the process of customizing an LLM in a way that makes it better at a given task than the original version. For example, a company building a product recommendation engine might wish to enhance a model’s ability to understand shoppers’ buying preferences.

Traditional fine-tuning methods update all of an LLM’s parameters, the configuration settings that determine how it processes data. Tinker uses a different approach known as Low-Rank Adaptation, or LoRA for short. It requires a fraction of the computing resources as standard fine-tuning and simplifies model deployment tasks.

LoRA extends the LLM being fine-tuned with a relatively small number of additional parameters. From there, it only fine-tunes those additional parameters instead of the entire model. That approach takes less time and doesn’t require as much infrastructure.

LoRA’s efficiency comes with certain tradeoffs. Notably, LLMs trained using the method often provide lower output quality than models fine-tuned the usual way. In September, Thinking Machines Labs detailed an implementation of LoRA that it says can provide similar output quality as standard fine-tuning methods.

Tinker simplifies several of the technical tasks involved in launching LoRA fine-tuning runs. Usually, developers have to optimize their training code to run on a large number of graphics cards, which requires highly specialized skills. Tinker skips that step. The service allows developers to implement a fine-tuning workflow as a simple, single-processor Python script and automatically adapts it to run on multiple graphics cards.

The service also eases other aspects of the fine-tuning process. Customers have access to a tool that can periodically save an LLM during a training run. If the model encounters an error, developers can revert it to the most recent reliable version instead of starting from scratch.

Tinker rolled out into general availability today with several new features that weren’t available before. The first addition is a so-called sampling tool that enables developers to send test prompts to an LLM while it’s being fine-tuned. From there, they can analyze the model’s responses to determine if the fine-tuning is effective. 

Thinking Machines also expanded the list of open-source LLMs that Tinker can train. 

The company has added support for Kimi K2, a model with one trillion parameters. It’s highly adept at tool use, a term for interactions with external applications. Many LLMs struggle to complete tasks that involve a large number of external applications. Kimi K2, in contrast, can carry out up to 300 tool invocations per task.

Kimi K2 is joined by support for two vision models called Qwen3-VL-30B-A3B-Instruct and Qwen3-VL-235B-A22B-Instruct. The former LLM features a lightweight design optimized for hardware efficiency, while the latter can take on more complex tasks. Qwen3-VL-235B-A22B-Instruct also has a larger context window.

The launch of Tinker comes amid reports that Thinking Machines is seeking to raise $5 billion in new funding at a $50 billion valuation. If the round closes, the company will almost certainly launch additional products to meet investors’ likely high revenue expectations. Thinking Machines may accelerate its products development roadmap by acquiring other AI tooling providers.

Photo: OpenAI