April 28, 2026
Stealth mode? More like roast mode
Laguna XS.2 and M.1
Poolside finally steps out of hiding, and the comments instantly turned into a benchmark brawl
TLDR: Poolside launched two new coding-focused AI models and, crucially, released the software layer that makes them act like working agents. The community liked the practical tooling, but the loudest debate was whether rival smaller models still look better on the scoreboard.
After months of “are they ever actually launching?” energy, Poolside has finally rolled out its first Laguna models: M.1, the big flagship, and XS.2, the smaller one it’s letting people download and use more freely. The company says these models are built for long, complicated coding jobs, and it also dropped the same agent runtime it uses internally — basically the software “driver” that helps the model act on tasks instead of just chatting. That part got a surprisingly loud round of applause.
But the real show was in the comments, where the community immediately split into two camps: “nice launch” and “okay, but why is a rival model still beating this?” One of the spiciest reactions came from users pointing out that Qwen’s smaller model appears to outperform not only Laguna XS.2, but sometimes even the much larger M.1. Translation for non-model nerds: people were basically saying, why is the smaller kid on the block outrunning the giant? That sparked instant side-eye over which competitors were included in the charts — and which were mysteriously absent.
Still, there was genuine praise. Early testers said Poolside’s “pool” agent feels fast and actually works well with outside tools, while others said shipping the real agent harness with the model is a big deal because most companies dump the brain and leave developers to build the body themselves. Even Poolside’s own designer jumped in to thank commenters and plug a clearer deeper-dive post, which only added to the vibe: stealth is over, and now the internet jury is very much in session.
Key Points
- •Poolside released Laguna M.1 and Laguna XS.2, the first two models in its Laguna family, along with its agent runtime in preview.
- •Laguna XS.2 is Poolside’s first open-weight model, with weights available under an Apache 2.0 license, while both models are temporarily free via API and OpenRouter.
- •Laguna M.1 is a 225B total-parameter, 23B activated MoE model trained from scratch on 30 trillion tokens using 6,144 NVIDIA Hopper GPUs.
- •Laguna XS.2 is a 33B total-parameter, 3B activated second-generation MoE model trained on 30 trillion tokens and positioned as a smaller agentic coding model.
- •Poolside reported benchmark results using the Harbor Framework and its ACP-based agent harness, including SWE-bench Pro scores of 46.9% for M.1 and 44.5% for XS.2.