Laguna-XS.2: Poolside's Open-Source Coding Model Release

Introduction

Poolside has officially announced the release of Laguna-XS.2, a groundbreaking coding model designed to redefine local software engineering capabilities. Released on April 28, 2026, this model marks a significant milestone in the open-weight space, bridging the gap between enterprise-grade performance and consumer hardware accessibility. Unlike previous iterations that required massive cloud clusters, Laguna-XS.2 is engineered for practical deployment, offering developers a powerful tool that can run directly on their local machines without compromising on reasoning capabilities. This release signifies a shift towards democratizing high-performance AI for individual engineers and small teams who need robust coding assistance without recurring infrastructure costs. The model represents a new standard for agentic coding, prioritizing efficiency and accessibility over raw parameter count.

This announcement is particularly timely as the industry moves towards more specialized models for specific engineering tasks. The combination of open-source licensing and local runnability sets a new precedent for how coding assistants are delivered to the end-user.

Key Features & Architecture

The architecture of Laguna-XS.2 is built on a 33-billion parameter Mixture-of-Experts (MoE) structure, which ensures high efficiency. Only 3 billion parameters are activated per token, drastically reducing memory footprint while maintaining model density. The model supports native reasoning, featuring interleaved thinking between tool calls to enhance complex task execution. It utilizes Sliding Window Attention with per-head gating in 30 of 40 layers, optimizing attention mechanisms for long-context tasks. Additionally, the KV cache is quantized to FP8, further reducing memory requirements per token. This efficiency makes the model compact enough to run locally on a Mac with 36 GB RAM, a feat previously unattainable with models of this scale. The open-source license under Apache 2.0 encourages widespread adoption and modification by the community.

The technical specifications focus heavily on reducing inference costs and memory usage without sacrificing intelligence. This allows developers to experiment with advanced coding tasks on consumer-grade hardware.

Performance & Benchmarks

Performance benchmarks highlight the model's strength in coding and reasoning tasks. It achieves 68.2% on SWE-bench Verified and 62.4% on SWE-bench Multilingual, demonstrating superior code generation and understanding. On SWE-bench Pro, it scores 44.5%, and on Terminal-Bench 2.0, it scores 30.1%. Trained on 30 trillion tokens using the Muon optimizer, the model ensures broad knowledge coverage. The 128K context window allows for handling entire codebases, while the 8K output token limit provides sufficient space for detailed responses. These metrics place it competitively against larger closed-source models in specific engineering domains.