Poolside's Laguna XS.2 lands open weights

This is a meaningful open-weights entry in the coding-model race: Poolside is not just shipping a checkpoint, it’s shipping the training story, agent harness, and deployment path around it.

• –Poolside claims XS.2 reaches 44.5% on SWE-bench Pro and 30.1% on Terminal-Bench 2.0, which puts it in the competitive bracket for 30B-class agentic models
• –The 33B A3B MoE design is the headline efficiency story: strong capability without the inference cost of much larger dense models
• –Apache 2.0 licensing matters here, because it makes the model materially more usable for commercial teams than many “open” releases with restrictive terms
• –The company’s own benchmark setup is an internal agent harness, so the scores are directionally useful but not apples-to-apples with every public leaderboard
• –Local support across vLLM, Transformers, TRT-LLM, Ollama, and MLX lowers adoption friction for developers who want to test or fine-tune it quickly