Poolside releases Laguna S 2.1, claims it's the West's most powerful open model

Poolside released Laguna S 2.1, an open-source 118-billion-parameter Mixture-of-Experts model for agentic programming. With 8 billion active parameters per token and a 1 million token context window, it achieves 70.2% on Terminal-Bench 2.1 and 59.4% on SWE-Bench Pro, matching or outperforming models like DeepSeek V4 lite, NVIDIA Nemotron 3 Ultra, and Thinking Machines Inkling, which are 2.5 to 8 times larger. The company published full benchmark trajectories for reproducibility. Training took under four weeks on 4,000 H200 GPUs, and Poolside plans releases every five weeks. Founder Jason Warner, former GitHub CTO, spent 18 months building a "model factory" for automated experimentation and reinforcement learning. The open-source market is currently dominated by Chinese models like Xiaomi, Tencent, Z-AI, and Moonshot's Kimi K3, the latter scoring 88.3% on Terminal-Bench 2.1 versus Laguna's 70.2%. U.S. policy discussions may restrict American use of Chinese open models, potentially benefiting Poolside.