Backed by NVIDIA, Poolside AI has launched its lightweight, local-ready Laguna XS 2.1 model alongside a flagship enterprise variant to challenge global coding dominance with permissive, open weights.
Poolside AI, a heavily funded US-based foundation model lab, released its first public AI models, Laguna XS 2.1, the lightweight model, and Laguna M.1, the heavyweight model, tailored specifically for agentic coding, on 2 July, 2026.
Laguna XS 2.1 is their fast, fully open-weight coding model. It features a 33-billion parameter Mixture-of-Experts (MoE) architecture but activates only 3 billion parameters per token. It is compact enough to run locally and privately on a single desktop, laptop, or MacBook GPU. It features a major performance jump on the SWE-bench Multilingual coding benchmark over the previous XS.2 iteration. Laguna M.1 is their flagship, and highly capable model optimized for long-horizon software engineering in enterprise and government settings. It boasts 225 billion total parameters with 23 billion active parameters via MoE.
The lightweight models are distributed with open weights under the OpenMDW-1.1 license, following an initial Apache 2.0 open-license model architecture for the community. Solo developers can run Laguna XS 2.1 locally for free through platforms like Ollama, vLLM, or llama.cpp with zero code leaving their machine. Poolside offers these models for free during a limited-time preview through its proprietary API and OpenRouter.
Poolside, which previously focused strictly on high-level enterprise and government clients, raised roughly $2 billion at a $12 billion valuation with major backing from Nvidia. The company explicitly released these open weights to challenge the market dominance of Chinese AI labs, like Alibaba’s Qwen and DeepSeek, which have historically led the open-weight coding assistant sector.
