Laguna S 2.1

Poolside's flagship open-weight coding model, released 2026-07-21 and pitched as 'the West's most capable open-weight model' -- a 118B-parameter (8B active) sparse MoE built for agentic software engineering. A scale-up of the Laguna XS family trained on the same pre-training data as Laguna XS 2.1, sitting between XS 2.1 (33B-A3B) and the proprietary Laguna M.1 (225B-A23B). 1M-token context, text-only, native interleaved 'thinking' (max-thinking on by default, toggleable per request). Open weights under OpenMDW-1.1 (Linux Foundation permissive model license); ships BF16/FP8/INT4/NVFP4 checkpoints plus official GGUF and MLX conversions and DFlash speculative-decoding draft models. Small enough to run on a single NVIDIA DGX Spark (BF16 checkpoint ~236 GB). Served first-party via platform.poolside.ai/chat.poolside.ai and on OpenRouter/Baseten/Vercel AI Gateway/NVIDIA NIM at $0.10/M input, $0.20/M output ($0.01/M cache-read); free endpoint caps context at 256K, paid dedicated gives the full 1M. Founded by Jason Warner (ex-GitHub CTO) and Eiso Kant. All benchmarks are Poolside self-reported (coding-only suite); no independent re-evaluation exists yet.