Poolside, the AI lab co-founded by former GitHub CTO Jason Warner, released its first public models on April 28: Laguna M.1 and Laguna XS.2. M.1 is a 225-billion-parameter Mixture-of-Experts model with 23 billion active per token, trained from scratch on 30 trillion tokens across 6,144 NVIDIA H200 GPUs. It scores 72.5% on SWE-bench Verified and 46.9% on SWE-bench Pro. XS.2 is far smaller: 33 billion total parameters, only 3 billion active, with 256 routed experts across 40 layers. It still hits 68.2% on SWE-bench Verified and 44.5% on SWE-bench Pro, and runs on a single GPU or a Mac with 36GB of RAM. XS.2 ships under Apache 2.0 on Hugging Face. Both models are free to use for a limited time via Poolside’s API and OpenRouter.
Poolside trained both using the Muon optimizer instead of AdamW, reaching the same loss in 15% fewer steps. Their RL loop is fully asynchronous: actors run coding tasks in sandboxed containers while the trainer pulls trajectories continuously, with weight transfers under 5 seconds via RDMA. They also released pool, a terminal coding agent implementing the Agent Client Protocol, and Shimmer, a cloud sandbox for building web apps. The company has raised over $626 million.
XS.2 scoring 44.5% on SWE-bench Pro with only 3 billion active parameters shows how efficiently MoE routing compresses coding ability into a small inference budget. GLM-5, another recent MoE release, made a similar case for sparse architectures. If coding skill keeps scaling with data and RL rather than raw parameter count, the cost of running capable software agents drops fast.
Sources:
- Introducing Laguna XS.2 and Laguna M.1 (Poolside Blog)
- Laguna XS.2 and M.1: A Deeper Dive (Poolside Blog)
- Laguna XS.2 Model Card (Hugging Face)
- pool: Poolside’s Coding Agent (GitHub)
- Poolside Founders Interview (Upstarts Media)
Disclaimer: For information only. Accuracy or completeness not guaranteed. Illegal use prohibited. Not professional advice or solicitation. Read more: /terms-of-service
Reuse
Citation
@misc{kabui2026,
author = {{Kabui, Charles}},
title = {Poolside {Laguna} {M.1} and {XS.2:} {A} {\$626M} {Startup’s}
{First} {Public} {Models} {Target} {Agentic} {Coding}},
date = {2026-05-04},
url = {https://toknow.ai/posts/poolside-laguna-m1-xs2-moe-agentic-coding-open-weights/},
langid = {en-GB}
}