Laguna XS 2.1 Released (5-minute read)

Poolside has unveiled Laguna XS 2.1, an upgrade to its existing XS.2 model. This model has a Mixture-of-Experts (MoE) architecture with a total of 33B parameters, using 3B active parameters per token, optimized for agentic coding and long-horizon work in local environments.

Key Improvements and Performance:

• The SWE-bench Multilingual score is 63.1%, up 5.4 points from the previous version.
• It shows stronger performance on Terminal-style tasks.
• It ships together with the DFlash speculator model, doubling inference speed (tok/s).

Deployment and License:

• It supports vLLM, SGLang, NVIDIA TensorRT-LLM, HF transformers, Ollama, with llama.cpp support planned soon.
• FP8, INT4, NVFP4 quantized checkpoints are provided for environments with limited VRAM and compute resources.
• To support the community's deployment of open models, the OpenMDW-1.1 license is applied, aiming for a more open model ecosystem.