Poolside Disrupts AI Landscape with Free, Open-Source Coding Model: Laguna XS.2

The AI world often feels like a high-stakes game of tennis, with giants like Anthropic and OpenAI constantly volleying new, powerful, and often expensive proprietary models. Meanwhile, companies in China have been pushing the boundaries with open-licensing and budget-friendly options. But what happens when a U.S. startup throws a curveball with an open-source model that’s both affordable and remarkably powerful? Enter Poolside, a San Francisco-based AI startup that’s shaking things up with its new Laguna large language models.

Poolside Unveils Laguna: Powering Agentic Workflows

Poolside has launched two impressive models: Laguna M.1 and Laguna XS.2. While M.1 is a proprietary powerhouse designed for complex enterprise and government tasks, it’s the XS.2 that’s generating the most buzz. This 33-billion parameter Mixture of Experts (MoE) model is released under the permissive Apache 2.0 license, making it accessible for community innovation and fine-tuning. The real game-changer? Laguna XS.2 is optimized for agentic workflows, meaning it can do more than just chat – it can write code, utilize third-party tools, and take autonomous actions.

What makes XS.2 particularly disruptive is its ability to run locally on a single GPU. Developers can download and execute this model on their personal computers without an internet connection, ensuring privacy and security. This is a significant departure from the cloud-dependent nature of many cutting-edge AI models.

Poolside isn't just stopping at the models. They've also introduced two complementary tools:

• pool: A terminal-based coding agent harness that acts as the same system Poolside uses for its internal reinforcement learning training.
• shimmer: A web-based, mobile-optimized environment for agentic coding development and interactive previews. Imagine writing and testing code with a powerful AI assistant right on your smartphone – that’s the vision shimmering offers.

Performance That Punches Above Its Weight

Don’t let the “smaller” moniker fool you; Laguna XS.2 is a performance beast. Benchmarks show it achieving remarkable results on coding tasks, even competing with much larger models. On the SWE-bench Pro, XS.2 scored an impressive 44.5%, nearly matching its larger sibling, M.1, which also performed admirably, nearing the capabilities of models like Qwen-3.5 and DeepSeek V4-Flash. On the SWE-bench Verified track, M.1 scored 72.5%, demonstrating its strength in complex, long-horizon software engineering problems.

These benchmarks highlight Poolside’s strategic focus on agentic reinforcement learning and meticulous data curation. Their approach allows models like XS.2 to punch well above their weight class, offering a compelling alternative for developers who value local execution and customizable workflows.

Running Laguna XS.2 Locally: What You Need

For developers eager to get their hands on Laguna XS.2, Poolside has provided clear hardware recommendations:

• Apple Silicon Macs: A minimum of 36 GB of unified memory is required, making M5 Max chip models a good starting point.
• PCs and Linux: With 4-bit quantization (Q4), the model can run on GPUs with 24 GB to 32 GB of VRAM, such as the NVIDIA RTX 5090. For more demanding tasks, higher VRAM options like the RTX PRO 6000 Blackwell are recommended.
• Storage: Plan for around 70 GB for the full model or 20-35 GB for a compressed version.

Poolside recommends using tools like Ollama or their own pool agent to manage the model’s capabilities seamlessly on consumer hardware.

The Power of Open Source and Agentic AI

Poolside’s decision to release Laguna XS.2 under the Apache 2.0 license is a strategic masterstroke. They firmly believe the West needs robust open-weight models, and community involvement is the fastest way to innovate and improve. By placing a powerful, agentic coding model in the hands of developers worldwide, Poolside is fostering a vibrant open-AI ecosystem.

Their core philosophy is that software development is the ultimate benchmark for general intelligence. The ability to plan, reason, and manipulate abstract systems is at the heart of human cognition, and Poolside’s AI agents are designed to embody these capabilities by writing and executing code. This represents a fundamental shift from AI simply using tools to AI actively building systems.

With its cutting-edge performance, commitment to open source, and innovative developer tools, Poolside’s Laguna release is setting a new standard for agentic AI in 2026. The future of work, it seems, is indeed agentic, and the language of that future is code.