Poolside releases Laguna XS 2.1
Laguna XS 2.1 is a 33B MoE with 3B active parameters, 63.1% on SWE-bench Multilingual, and open weights deployable on Ollama, vLLM, and HF Transformers.
Poolside released Laguna XS 2.1 on July 2, a 33-billion-parameter mixture-of-experts model that activates 3 billion parameters per token. On SWE-bench Multilingual it scores 63.1 percent, up 5.4 points from the previous XS.2 release. Context window is 256,000 tokens.
Pricing on the hosted API is $0.10 per million input tokens, $0.20 per million output tokens, and $0.05 per million cache-read tokens. Weights are released under the OpenMDW-1.1 license. Local deployment is supported on Ollama, vLLM, SGLang, TensorRT-LLM, and Hugging Face Transformers, with llama.cpp support listed as coming. Quantized versions at FP8, INT4, and NVFP4 precision are available.
The model targets multilingual code work specifically, so the SWE-bench Multilingual number matters more than the standard English-only leaderboard position. Whether a 5.4-point improvement justifies a switch depends on your task mix. How to judge an AI performance claim covers translating benchmark movement into something measurable in your own codebase.
The 256K context and low active parameter count are the infrastructure figures to notice. At 3B active parameters, inference cost on local hardware stays low relative to the context window, which helps for long-file or multi-file code tasks. If you are evaluating whether to run open models at all, open models good enough for operators covers when the local-deployment tradeoff clears the bar.