Laguna S 2.1

Poolside's new mid-tier agentic-coding model, released 21 July 2026 to sit between Laguna XS 2.1 (33B-A3B) and the flagship Laguna M.1 (225B-A23B). A 118B-total / 8B-active Mixture-of-Experts transformer: 48 layers (12 global-attention, 36 sliding-window at 512 tokens), 256 routed experts (top-10) plus one shared, and grouped-query attention (8 KV heads). Its headline feature is a 1M-token context window (1,048,576) — a 4× jump over the 256K of XS.2 and M.1. Released under OpenMDW-1.1 with base and post-trained checkpoints, FP8/NVFP4/INT4/GGUF quantizations, and a DFlash speculative-decoding draft model; serves on vLLM, SGLang, Transformers, TRT-LLM, and llama.cpp (BF16 weights ≈236GB, multi-GPU).

Reported benchmarks (as of 21 July 2026): Terminal-Bench 2.1 70.2, SWE-bench Multilingual 78.5, SWE-bench Pro (public) 59.4, DeepSWE 40.4, SWE Atlas (codebase QnA) 46.2, Toolathlon Verified 49.7. Unlike the fully self-reported figures on the earlier Laguna models, several of these are now third-party numbers — Terminal-Bench 2.1 and DeepSWE via Artificial Analysis, SWE Atlas via Scale AI's leaderboard, and Toolathlon via its official leaderboard — though Poolside remains unscored on the AA Intelligence Index.