Poolside AI just released two agentic coding models that punch above their weight class. Laguna M.1 is a 225B parameter Mixture-of-Experts model with 23B active parameters. Laguna XS.2 is a much smaller 33B total, 3B active. The bigger deal: XS.2 is Poolside's first open-weight release, available under Apache 2.0. Both models were trained from scratch on 30 trillion tokens using 6,144 NVIDIA Hopper GPUs.

The benchmarks tell an interesting story. Laguna M.1 hits 46.9% on SWE-bench Pro. XS.2, despite being dramatically smaller, reaches 44.5% on the same benchmark. That's solid for a 3B active parameter model, beating Gemma 4 (35.7%) and edging Devstral Small 2 on Verified tasks. But the Hacker News crowd noticed that Qwen3.6, another small MoE model at 35B total with 3B active, scores 49.5% on SWE-bench Pro and 51.5% on Terminal-Bench 2.0. That beats both Laguna models despite being a fraction of M.1's size. Poolside's comparison tables conveniently left Qwen3.6 out of the M.1 row, only including it alongside XS.2.

What makes this release different is Poolside's thesis about agents. They argue that tool calling is a transitional pattern. The real interface for agents should be writing and executing code directly.

Software is more flexible than predefined action sets. An agent that writes code can compose new behaviors on the fly and build ad-hoc systems for tasks no one predicted.

They're also releasing their Agent Client Protocol server, the same framework used for agent training and evaluation. Early testers report it integrates cleanly with editors like Zed.

The company background matters here. Co-founded by Jason Warner (former GitHub CTO) and Eiso Kant, Poolside spent its early years in stealth serving government and defense clients who need air-gapped, high-security deployments. A $500 million Series B led by Bain Capital Ventures bankrolled the GPU cluster. Now they're taking that government-grade infrastructure into the open ecosystem, arguing the West needs strong open-weight models. Whether XS.2 can carve out space against Qwen's increasingly dominant small models is an open question. But Apache 2.0 licensing and built-in agent tooling give developers a real alternative.