Here is a streamlined, engaging rewrite of the announcement.
Poolside Disrupts the AI Race with Powerful, Open-Weight “Laguna” Models
The current AI landscape feels like a high-stakes tennis match. In one corner, incumbents like Anthropic and OpenAI volley increasingly expensive, proprietary models back and forth. In the other, efficient, open-licensed competitors from abroad are democratizing access to frontier-level intelligence. Today, San Francisco-based startup Poolside entered the court with a surprise: a new, highly performant, and U.S.-developed pair of Large Language Models (LLMs) that prioritize agentic workflows—AI that does more than chat; it writes code, uses tools, and executes autonomous tasks.
Meet the Laguna Family
Poolside has unveiled two flagship models, both trained from scratch to master long-horizon software engineering: Laguna M.1 (Proprietary / Enterprise): A massive 225B parameter Mixture of Experts (MoE) model. Built for high-consequence government and enterprise environments, it excels at complex reasoning and multi-step planning. It is currently available via API and platforms like OpenRouter, Ollama, and Baseten. Laguna XS.2 (Apache 2.0 Open Source): A versatile 33B parameter MoE model. Engineered for efficiency, this model is a powerhouse for local development. Developers can run it entirely offline on a single GPU, ensuring maximum privacy and fine-tuning control.
Beyond Chat: The “Pool” Ecosystem
What makes these models special isn’t just their training; it’s how you use them. Poolside has also launched: “pool”: A terminal-based coding agent that allows developers to integrate AI agents directly into their local workflows. “shimmer”: A cloud-native, mobile-optimized IDE that lets developers build and test software in an instant-on virtual environment—even from a smartphone.
Why Poolside Wins on Performance
Poolside’s secret sauce lies in its “Model Factory.” By using a custom Muon optimizer (which accelerates training by 15%) and the AutoMixer system—which selectively curates trillions of tokens of data—they have created models that “punch up.” Despite being smaller, the Laguna XS.2 achieves a 44.5% score on the difficult SWE-bench Pro, outperforming much larger models like Gemma 4 or Claude Haiku 4.5. The larger M.1 model secures a 72.5% on the Verified track, putting it in the same league as the industry’s top-tier models.
Running Laguna Locally
Because the XS.2 model is optimized for 4-bit quantization, you don’t need a supercomputer to run it. Mac Users: Apple Silicon with 36GB of unified memory is the baseline. PC Users: You’ll need a modern GPU (like an RTX 5090 or 4090) with at least 24GB of VRAM to handle the model’s reasoning capabilities effectively.
The Big Picture: A Commitment to Open Weights
By releasing Laguna XS.2 under an Apache 2.0 license, Poolside is taking a firm stance: they believe the West needs strong, accessible, open-weight models to compete globally. Unlike labs that rely on fine-tuning Chinese base models, Poolside’s “from-scratch” approach demonstrates a long-term commitment to sovereign, secure, and reproducible AI. For developers, researchers, and enterprise architects, the message is clear: the future of work is agentic, and with Poolside, that future is now accessible. Ready to start building? Check out the models on [Hugging Face] or dive into their technical blog.