Laguna XS 2.1 is poolside’s second-gen 3B-active MoE explicitly designed for agentic coding on local hardware — 33B total params, 262K context, FP8 KV cache, mixed SWA/global attention (3:1 across 40 layers), native reasoning with interleaved thinking between tool calls. Beats its predecessor Laguna XS.2 by +5.4pp on SWE-bench Multilingual (63.1% vs 57.7%) and leads all open-weight models <70B on SWE-bench Verified (70.9%) per the lab’s Harbor Framework runs (mean pass@1 over 4 attempts). Launch-day support in vLLM, SGLang, Transformers, TRT-LLM, llama.cpp, Ollama. OpenMDW-1.1 license (commercial-friendly). The DFlash 5-layer speculator (~70% acceptance on coding) is available for speculative decoding. This is the strongest open-weight agentic coding model in the 1B–70B active-param range for local deployment.