Poolside Launches Two New Agentic Coding Models, Including Open-Weight Release, Free via API for Limited Time

Poolside launches two powerful agentic coding models — Laguna M.1 (225B) and open-weight Laguna XS.2 (33B) — both free via API for a limited time, delivering top-tier SWE-bench Pro scores and built for complex, long-horizon coding tasks using reinforcement learning trained on 30 trillion tokens.

• Poolside releases two new agentic coding models, Laguna M.1 (225B parameters) and Laguna XS.2 (33B parameters), with Laguna XS.2 available as an open-weight model under Apache 2.0 license, both free to use for a limited time via API and OpenRouter.
• Laguna M.1 achieves 46.9% on SWE-bench Pro and 40.7% on Terminal-Bench 2.0, while the smaller Laguna XS.2 reaches 44.5% on SWE-bench Pro, with both models trained from scratch on 30 trillion tokens using NVIDIA Hopper GPUs and the Muon optimizer for improved training efficiency.
• Both models are built for long-horizon agentic tasks, featuring a fully asynchronous online reinforcement learning system and an Agent Client Protocol harness that enables agents to write and execute code, compose actions, and build ad-hoc systems rather than being limited to fixed tool-calling interfaces.