Together Computer Inc., which does business as Together AI, today said that it has acquired CodeSandbox B.V, a startup that builds isolated cloud-based code execution environments known as devboxes.

Together AI said combining the two companies will enable it to add a code interpreter component to its Together Inference Platform. The goal is to enable developers to run inferencing on open-source models natively for scenarios such as agentic workflows, data analysis, dynamic file processing, custom visualizations and artificial intelligence-assisted development.

CodeSandbox is an online development platform for collaborative application development in the cloud. It enables developers to write, run and share code from a web browser without local setup. It supports various programming languages and frameworks and claims a base of more than 4.5 million monthly users.

Founded in 2022, Together AI develops an open-source infrastructure platform, models and services that organizations can access via an application programming interface to train, fine-tune and run AI models. It touts lower cost and up to four times better performance than services from hyperscale cloud vendors. The company has raised $288 million in venture funding, including a $106 million round in March that valued it at $1.25 billion.

“We describe [our platform] as an AI acceleration cloud,” said Chief Executive Vipul Parakash. “We are a focused, pure-play cloud service that does a lot of fundamental research on how to make generative AI workloads highly performant, highly scalable and economically rational.”

As part of the acquisition, it’s providing early access to what it says is unique code interpretation technology via an API that allows developers to execute code at large scale. That’s faster than submitted code to a compilation process and also allows the to leverage more complex logic in generative AI applications.

Faster than the hyperscalers

Together AI says its platform delivers four times the end-to-end performance of other cloud options while supporting more than 100 large language models. “We work at all layers,” Parakash said. “We find the best possible kernels and write the most performant kernels” for the particular graphics processing unit architecture, such as Nvidia Inc.’s Hopper.

The company is also developed various techniques for optimization like speculative decoding, which uses small models to improve the efficiency of text predictions in large language models. “We are building all of this as a very easy-to-use end-to-end platform,” Prakash said.

With developers increasingly using generative AI to create code, he said, a sandbox reduces the risks of submitting code for compilation and testing outside the development platform.

“The Code Sandbox platform is a way to make models actionable,” he said. “In milliseconds, you can fire up a temporary virtual machine that compiles the code, runs it and returns the output.” The platform is optimized for JavaScript and Python but can be used with any language.

Customers can sign up here for early access to the sandbox technology and here for a beta test version of the software development kit.

Photo: Unsplash