Laguna XS 2.1

Poolside's compact open-weight coding model, released 2026-07-02 -- a 33B-parameter (3B active) sparse MoE for agentic software engineering, and the smaller sibling that Laguna S 2.1 scales up from (same pre-training data). An upgrade of Laguna XS.2 with a +5.4-point jump on SWE-bench Multilingual (63.1% vs 57.7%) and a license move from Apache-2.0 to OpenMDW-1.1. 256K-token context, text-only, native interleaved reasoning (toggleable per request). Ships BF16/FP8/INT4/NVFP4 checkpoints plus a DFlash 0.5B speculator; compact enough to run on a Mac with ~36 GB unified memory (BF16 download ~67 GB). Served first-party via platform.poolside.ai and on OpenRouter/Ollama/NVIDIA NIM at $0.10/M input, $0.20/M output ($0.05/M cache-read), 256K context. All benchmarks are Poolside self-reported (coding-only suite); no independent re-evaluation exists yet.