Poolside develops specialized code generation models

Poolside utilizes 10,000 Nvidia GPUs to train its Laguna model family, which features Mixture of Experts architecture. The startup targets Global 2000 companies by offering on-premises deployment to ensure data sovereignty and security for complex coding workflows.

Poolside develops specialized code generation models

Jason Warner and Eiso Kant founded Poolside in April 2023. Warner and Kant met in 2017 and spent six years planning an AI-driven assistive tool suite for developers. Warner previously directed engineering at GitHub and worked at Canonical and Heroku. He also incubated the GitHub Copilot tool. Kant previously co-founded the engineering analytics firm Athenian. The startup moved its headquarters from the United States to Paris. In August 2023, the company raised $126 million in a seed round. This round included participation from Redpoint Ventures, Bpifrance, NewWave, Xavier Niel, Felicis Ventures, Motier Ventures, and FRST. The company recently secured $500 million in Series B funding led by Bain Capital Ventures. Other participants in this round include eBay through eBay Ventures and Nvidia. Additional investors include LG Technology Ventures, Felicis Ventures, Redpoint Ventures, Citi Ventures, Capital One Ventures, HSBC Ventures, DST Global, Stepstone Group, Schroders Capital, Premji Invest, Dorsal Capital, BAM Elevate, Adams Street, and Fin Capital. This capital brings the total amount raised to $626 million. While the valuation currently sits at $3 billion, the company uses 10,000 Nvidia GPUs to train future models. I view this massive accumulation of capital as a move to build a foundation of compute power that most software development startups cannot match.

The Laguna model family

The Laguna model series uses Mixture of Experts architecture for code generation. Each model includes native reasoning capabilities that users can toggle on or off per request. These models are text-to-text systems. Poolside trains these models by completing millions of tasks in tens of thousands of real world software projects. By engineering proprietary techniques that allow models to emulate the cognitive and adaptive capabilities of elite developers, the company attempts to streamline complex coding workflows for enterprise clients who require strict control over their proprietary software logic. The Laguna XS 2.1 model works well for fast agentic coding and rapid iteration. For more demanding needs, the Laguna M.1 model handles complex, multi-step coding tasks where agent quality matters more than raw throughput. The Laguna S 2.1 model is the preferred option for long-horizon agentic coding and extended tool-use workflows.

Model Total Parameters Active Parameters Context Window GPU Memory (FP8)
Laguna S 2.1 118B 8B 1M tokens 192 GB
Laguna XS 2.1 33B 3B 256K tokens 96 GB
Laguna M.1 225B 23B 256K tokens 384 GB

Deployment and market competition

You already know that GitHub Copilot has over 1.8 million paying users and 50,000 business customers, but Poolside targets a different segment. The company focuses on Global 2000 companies and public-sector agencies. These clients prioritize data sovereignty. Poolside allows organizations to deploy models directly within their own secure environments. This on-premises deployment capability helps mitigate security risks. I find this focus on local environments a logical response to the privacy concerns surrounding cloud-only tools. The AI coding tools market could reach a value of $27 billion by 2032 according to Polaris Research. Many startups compete in this space. Magic raised $320 million in August. Codeium closed a $150 million fundraising round. Cognition secured $175 million at a $2 billion valuation.

GitHub Copilot uses OpenAI’s Codex model to provide suggestions. Amazon CodeWhisperer helps developers in e-commerce, healthcare, and finance. It supports languages such as Python, Java, JavaScript, TypeScript, and C++. CodeWhisperer uses machine learning algorithms trained on Amazon’s internal codebase and a selection of open-source code. It also includes a built-in security scan feature to identify potential vulnerabilities. The tool provides code suggestions directly in integrated development environments, which can help reduce coding errors and improve productivity. Developers can customize the tool according to specific coding styles and preferences. GitHub Copilot also integrates with popular IDEs like JetBrains, Neovim, and Visual Studio Code, using the Codex model to examine code and comments to provide recommendations. Poolside has not launched its first product yet. Will developers adopt a tool that has no public release?

airtrain.ai
airtrain.ai

The airtrain.ai newsroom covers AI research, models and the tools built on them.

More on this topic

Stay ahead of AI

Get the week's most important AI stories delivered to your inbox every Monday.

No spam. Unsubscribe anytime.

More Stories