Poolside

What Poolside is

Poolside is an AI company that builds models for software work. On 28 April 2026 it released two of them, Laguna M.1 and Laguna XS.2, together with pool, the agent runtime that its reinforcement learning runs on ^[1]. Until then, the company said, its work had concentrated on government and public-sector customers that need air-gapped deployment, meaning systems that run on machines cut off from outside networks ^[1].

The two Laguna models

Laguna M.1 is the larger of the pair. It has 225 billion parameters in total and activates 23 billion of them per token, meaning the model holds far more weights than it uses for any single step of computation. Poolside trained it from scratch on 30 trillion tokens, using 6,144 Nvidia Hopper GPUs ^[1].

Laguna XS.2 is much smaller: 33 billion parameters in total with 3 billion active per token, released under the Apache 2.0 licence as the company's first open-weight model ^[1].

On SWE-bench Verified, a benchmark that measures whether a model can resolve real software issues, Laguna M.1 resolves 72.5 percent of tasks and Laguna XS.2 resolves 68.2 percent ^[1]. The gap between the two is small relative to the difference in size, which is the sort of result that makes a small model interesting to people who have to pay for inference.

Code as the interface

The company makes a broader argument about how agents should talk to software. A fixed set of tool-calling APIs, in its view, limits what an agent can do to the actions someone anticipated in advance. Code is a more expressive interface, because an agent that writes and runs code can compose operations rather than pick from a menu ^[1]. That position shapes the models: they are trained for software work rather than for calling a fixed catalogue of functions, and pool, the runtime, is the environment where that behaviour is learned and executed ^[1].

Open weights and access

The release arrived in stages. At launch the M.1 weights were not open; Poolside listed open weights as coming and said access could be requested ^[1]. The XS.2 weights reached Hugging Face on 23 April 2026, and the M.1 weights followed on 15 June 2026 ^[2]. For a limited time both models were free through Poolside's API and through OpenRouter ^[1], which let developers try them without a commitment.

A small research group

The models come out of Poolside's Applied Research organisation, which has about 60 people ^[1]. That is small for work at this scale, and it is part of the company's story: a compact group training models from scratch on thousands of GPUs, aiming at customers who cannot send their code to a hosted service.

Poolside's path is a reminder that frontier model work is not only done by the largest labs. A company can start with a narrow, demanding customer base, namely organisations that need their models to run on isolated hardware, and use that constraint to justify building its own training stack, its own runtime and its own models, then open part of the stack up when it wants a wider audience ^[1][2].