Poolside releases the open-weight Laguna S 2.1 coding model

Sources: Poolside blog, HN discussion
Summary: Poolside released Laguna S 2.1 on 2026-07-21, a 118B-parameter Mixture-of-Experts model with 8B active parameters per token aimed at agentic coding, with open weights under the OpenMDW-1.1 license on Hugging Face in BF16, FP8, INT4, and NVFP4 formats and up to a 1M-token context. Poolside reports Terminal-Bench 2.1 70.2 percent, SWE-Bench Multilingual 78.5 percent, and SWE-Bench Pro 59.4 percent as mean pass@1 averaged over three to four attempts per task, says it is the first model trained with FP8-precision reinforcement learning, and states it was built in nine weeks. OpenRouter lists a free 256K-context tier and a paid endpoint at 0.10 USD per 1M input and 0.20 USD per 1M output tokens. Benchmark figures are vendor-reported and unverified.
Why it matters: A US lab shipping a low-cost open-weight coding model is a counterpoint to the day's narrative that American AI is uniformly closed, and the listed pricing undercuts proprietary coding APIs.