Laguna XS 2.1

Poolside's Laguna XS 2.1 is a 33B-parameter Mixture-of-Experts LLM specialized for agentic coding and long-horizon software-engineering work. It routes 8 of 256 experts (plus a shared expert) per token — roughly 3B active parameters — but keeps all experts resident, so its memory footprint equals the full quantized weight file (Q4_K_M is 18.9 GiB, the smallest published GGUF, making it a 24 GB-class card model). It supports a 262,144-token (256K) context via YaRN scaling for whole-repository tasks. Vendor evals report SWE-bench Verified 70.9%, SWE-bench Multilingual 63.1%, SWE-Bench Pro (public) 47.6%, and Terminal-Bench 2.0 37.5%. Released under the permissive OpenMDW-1.1 license, with first-party GGUF quants (Q4_K_M, BF16) published by the same poolside org. Runs as a standard GGUF coding LLM under llama.cpp or Ollama, driven by an agent client (OpenHands / Aider) over an OpenAI-compatible API.