Poolside AI debuts Laguna M.1 and XS.2

Coding models are getting split into two lanes. One lane chases raw benchmark muscle. The other chases something more practical — models that can actuall(poolside.ai), and maybe even work on local hardware. Poolside’s April 28 release lands squarely in that second lane. The company put out two new models, (poolside.ai)uilt around agentic coding workflows. (poolside.ai) ### What actually shipped? Poolside released(poolside.ai)aguna XS.2 as open weights on Hugging Face. It also tied both models to two preview experiences — a terminal agent and a broader interface called Shimmer — instead of treating them like plain chatbots. That framing matters because Poolside is selling a workflow, not just a checkpoint. (poolside.ai) ### What are these models? Both are mixture-of-experts models with 128K context windows, but they sit in very different si(poolside.ai)vated parameters per token, built for “long-horizon work on a local machine.” Laguna M.1 is the heavyweight foundation model in the family at 225B total parameters with 23B active. Basically, XS.2 is the lighter workhorse; M.1 is the bigger brain behind the lineup. (poolside.ai) ### Why does “active parameters” matter? Because MoE models(poolside.ai)through a smaller subset of experts. That means a model can have a large total parameter count but a much smaller active footprint at inference time. The result is the tradeoff Poolside wants to emphasize — more capability than a tiny dense model, but lower latency and compute than a same-size fully active model. XS.2’s 3B active figure is the clearest signal here. (huggingface.co) ### Why is XS.2 the more(poolside.ai)del. M.1 is important inside Poolside’s own stack, and the company says it finished pretraining at the end of last year. But XS.2 is the model outsiders can actually download, inspect, quantize, and try in their own setups. There is already an INT4 version on Hugging Face, which tells you Poolside expects people to care about memory footprint and deployment, not just leaderboard screenshots. (poolside.ai) ### What is Poolside optimizing (huggingface.co) describing both models as strongest when they can explore a codebase, edit files, run tests, and iterate on fixes. That is a different target from “answer this one question well.” It is closer to giving a model a terminal, a repo, and enough persistence to keep working through a task. (docs.poolside.ai) ### So is this a pure open-source play? Not really. Turns out this is more of a hybrid move. Poolside opened XS.2’s weights, but(poolside.ai)prise pitch. The company is still very focused on controlled deployment, governance, and use inside customer infrastructure. Open weights widen the funnel; the business still points toward paid agent systems. (poolside.ai) ### What should readers take away? The headline is not just “two new models.” It is that Poolside is trying to (docs.poolside.ai)one large internal anchor model and one lighter open-weight model meant to move faster in the wild. If that works, the interesting competition will not just be who has the smartest model — it will be who has the most usable coding loop. (poolside.ai)