The generative AI landscape keeps shifting—today, top labs race to deploy proprietary models with ever-increasing power, while others focus on accessibility. But few announcements carry the same disruptive potential as Poolside’s latest release: a free, open-source AI model designed to run coding agents on a single GPU, even offline.
Poolside, a San Francisco-based AI startup founded in 2023, has just unveiled two new large language models (LLMs) under its Laguna family. The standout is Laguna XS.2, a 33-billion-parameter Mixture of Experts (MoE) model licensed under Apache 2.0. Unlike its heavier counterpart, Laguna M.1, XS.2 is engineered for efficiency, enabling developers to fine-tune, quantize, and deploy agents locally without an internet connection. This makes it ideal for privacy-conscious use cases, from personal coding projects to enterprise workflows where data security is critical.
But the bigger surprise? Poolside is offering both models—M.1 and XS.2—at no cost through its API and partner platforms like OpenRouter, Ollama, and Baseten. While M.1 remains proprietary, its free availability signals a shift toward democratizing high-performance AI tools, particularly for teams that need agentic capabilities without vendor lock-in.
Why Laguna XS.2 stands out in the crowded AI model market
Most open-source models today rely on fine-tuning existing architectures, often borrowing from established systems like Alibaba’s Qwen series. Poolside took a different approach: it trained both Laguna models from scratch, ensuring a clean slate for optimization. The result is a model that balances performance with resource efficiency.
Key highlights of Laguna XS.2 include:
- 33-billion parameters (with just 3 billion active), making it lightweight enough for local deployment on a single GPU.
- Apache 2.0 license, allowing unrestricted commercial and non-commercial use.
- Agentic workflow support, enabling the model to write code, interact with tools, and execute tasks autonomously.
- Offline compatibility, ideal for air-gapped environments or teams prioritizing data privacy.
Poolside’s decision to go open-source wasn’t taken lightly. The company has spent years serving government and public-sector clients, where deployment speed and security are paramount. In a recent exchange on social media, Poolside’s post-training engineer George Grigorev explained why some agencies might prefer Poolside’s models over larger proprietary alternatives: “We can be faster to deploy, and we can literally ship weights in fully isolated environments on-prem, which might be critical for government/public sectors.”
How Poolside built Laguna XS.2 for speed and efficiency
Training a model from scratch at scale is no small feat. Poolside’s secret weapon is its proprietary Model Factory, an end-to-end training ecosystem. At its core lies Titan, a custom software platform designed to streamline the training process. To accelerate learning, the team employs Muon, an optimization tool that boosts training speed by roughly 15% compared to standard methods.
Muon works by ensuring each update to the model’s weights is mathematically balanced, preventing the AI from getting stuck or distracted during training. This is especially important when working with Poolside’s colossal training dataset—30 trillion tokens—curated using AutoMixer, a system that blends code, mathematical reasoning, and general web data to maximize the model’s problem-solving abilities.
Unlike many competitors that rely on data scraped indiscriminately from the internet, Poolside’s AutoMixer uses a swarm of 60 proxy models to evaluate different data mixes. The goal isn’t just volume; it’s precision. By identifying the optimal combination of inputs, the system ensures the final model excels at complex, long-horizon tasks like autonomous coding.
Beyond models: Poolside’s app ecosystem for agentic coding
Laguna XS.2 isn’t just another model to download and forget. Poolside has built a companion ecosystem to make agentic coding more accessible:
- pool: A coding agent harness that simplifies the deployment of Laguna models for autonomous workflows.
- shimmer: A mobile-optimized web environment where developers can write, test, and preview code using Laguna models on the go. The interface is designed for real-time collaboration, making it easier to iterate without switching between tools.
For teams that need a balance between power and flexibility, Laguna XS.2 offers a compelling alternative to cloud-only solutions. Whether you’re a solo developer prototyping a new tool or an enterprise team building internal agents, the model’s open licensing and local deployment capabilities remove barriers that have long plagued the AI industry.
What’s next for Poolside and Laguna XS.2?
The release of Laguna XS.2 marks a bold step toward making high-performance AI more accessible and secure. But Poolside isn’t stopping there. The company is already hinting at broader plans to expand its open-source offerings, with a focus on refining agentic capabilities and improving deployment options for developers.
For now, the focus is on community adoption. With Laguna XS.2 available today on Hugging Face, developers can start experimenting immediately. And as more teams integrate the model into their workflows, we may see a ripple effect—challenging the dominance of proprietary solutions while proving that innovation doesn’t always require massive budgets or black-box systems.
AI summary
Amerika merkezli AI startup Poolside, yerel ajan kodlama için tasarlanmış ücretsiz ve açık kaynaklı Laguna XS.2 modelini tanıttı. Laguna XS.2’nin özellikleri ve avantajları hakkında detaylı bilgi edinin.
