33B total / 3B active per token (MoE)
256K tokens
OpenMDW-1.1
Official small open-weight coding and long-horizon agent release from Poolside, positioned as a local-machine companion model that upgrades Laguna XS.2 with stronger multilingual software engineering and terminal-style behavior. The Hugging Face card and release blog describe Laguna XS 2.1 as a 33B total-parameter Mixture-of-Experts model with 3B activated parameters per token, using the same token-choice routing setup as XS.2 but a revised architecture built for lower KV-cache pressure and faster inference: grouped-query attention, 40 layers, and sliding-window attention with per-head gating in 30 of those layers. Poolside frames XS 2.1 as the efficient member of the July 2026 Laguna refresh, released alongside Laguna S 2.1, with broad public formats including BF16, FP8, NVFP4, INT4, and GGUF for local deployment. Public benchmark tables emphasize improved coding-agent performance over XS.2 and competitive results for its size, with Poolside reporting 35.0 on Terminal-Bench 2.1, 73.6 on SWE-Bench Multilingual, 42.4 on SWE-Bench Pro Public, 28.5 on DeepSWE, 38.5 on SWE Atlas Codebase QnA, and 32.7 on Toolathlon Verified. The release blog highlights a +5.4 point jump on SWE-Bench Multilingual relative to XS.2, while Poolside’s docs position the model for Poolside Agent CLI workflows and OpenAI-compatible integrations. Overall, the release is pitched as a compact, agent-oriented open model that preserves Laguna’s long-horizon coding identity while making deployment more practical on smaller hardware.