American AI startup Poolside launches free, high-performing open model Laguna XS.2 for local agentic coding

```json { "title": "Poolside Launches Free Open AI Coding Model Laguna XS.2", "metaDescription": "American AI startup Poolside releases Laguna XS.2, a free open-weight agentic coding model that runs locally on a Mac and outperforms Claude Haiku 4.5.", "content": "<h2>Poolside Launches Free, Open-Weight Laguna XS.2 — An Agentic Coding Model That Punches Well Above Its Weight</h2>\n\n<p>On April 28, 2026, American AI startup Poolside made its first-ever public model release, unveiling two new foundation models built specifically for agentic software development: <strong>Laguna M.1</strong> and <strong>Laguna XS.2</strong>. The headline act is Laguna XS.2 — a compact, open-weight model released under an Apache 2.0 license that is free to download on Hugging Face, capable of running locally on a Mac with 36 GB of RAM, and, remarkably, competitive with or superior to several larger proprietary models on key software engineering benchmarks. For developers, enterprises, and anyone watching the open-source AI coding race, this launch is a significant data point.</p>\n\n<h2>What Poolside Actually Released: Two Models, Two Products, One Big Debut</h2>\n\n<p>Poolside shipped two distinct models on launch day. <strong>Laguna M.1</strong> is a Mixture-of-Experts (MoE) architecture with 225 billion total parameters and 23 billion active parameters per token — a large, enterprise-grade model positioned at the higher end of the capability spectrum. <strong>Laguna XS.2</strong> is the open-weight counterpart: a 33 billion total parameter MoE with just 3 billion active parameters per token, designed to be efficient enough for local deployment while remaining capable enough for real agentic coding workflows.</p>\n\n<p>Both models were trained entirely from scratch on more than 30 trillion tokens using Poolside's proprietary Titan training codebase, the company's own data pipeline, and its own agent reinforcement learning (RL) infrastructure. According to Poolside's technical blog, all training was conducted on NVIDIA hardware, and Laguna XS.2 ships with Day 1 support for NVIDIA TensorRT-LLM, including an NVFP4 version optimized for NVIDIA's Blackwell architecture.</p>\n\n<p>Laguna XS.2's architecture includes Sliding Window Attention with per-head gating in 30 out of 40 layers, a 128K context window, and support for up to 8K output tokens. It is quantized to fp8 on OpenRouter for fast, cost-efficient agentic coding workflows. The model is supported across vLLM, Hugging Face Transformers, TRT-LLM, Ollama, and the mlx-lm framework — covering the most common local and cloud inference stacks. Notably, Poolside began pre-training XS.2 just five weeks before its public release date, and it is being shipped fully post-trained.</p>\n\n<p>Alongside the models, Poolside also released two products in preview: <strong>pool</strong>, a terminal-based coding agent and Agent Client Protocol client-server, and <strong>shimmer</strong>, a cloud development environment for iterating on web apps, APIs, and command-line tools. Both Laguna models are free to use for a limited time via Poolside's API and through OpenRouter, and are distributed through partners including Hugging Face, OpenRouter, Baseten, and Ollama.</p>\n\n<h2>Benchmark Performance: A 3B Active-Parameter Model That Beats Claude Haiku 4.5</h2>\n\n<p>The benchmark numbers are what make this launch genuinely newsworthy. Laguna XS.2 achieves a <strong>44.5% score on SWE-bench Pro</strong> — one of the most demanding real-world software engineering benchmarks available — and a <strong>30.1% score on Terminal-Bench 2.0</strong>, despite activating only 3 billion parameters per token during inference.</p>\n\n<p>To put those numbers in context: on SWE-bench Pro, Laguna XS.2 surpasses Anthropic's Claude Haiku 4.5, which scored 39.5%, and Google's Gemma 4 31B dense model, which scored 35.7%. Both of those are well-resourced models from two of the largest AI organizations in the world. Laguna XS.2 outperforms both while being small enough to run on consumer-grade Apple Silicon hardware.</p>\n\n<p>This efficiency is central to Poolside's design philosophy. The MoE architecture means the model activates only a fraction of its total parameters for any given token — keeping inference compute low without sacrificing task-relevant capability. For developers running agentic coding workflows locally, or organizations looking to minimize cloud inference costs, this tradeoff is meaningful in practice, not just on paper.</p>\n\n<p>The model's performance is also a product of Poolside's training methodology. The company developed an approach called <strong>Reinforcement Learning from Code Execution Feedback (RLCEF)</strong>, where models learn from actually executing code across 130,000 real-world code repositories. Rather than relying solely on human preference data or static benchmarks, the models are evaluated against real execution outcomes — a signal that is harder to game and more directly tied to practical software engineering ability.</p>\n\n<h2>Why This Matters: An American Company Playing the Open-Weight Game</h2>\n\n<p>Until today, the dominant narrative in open-weight AI has been largely a story about Chinese companies. DeepSeek's open model releases earlier in 2025 and 2026 reset expectations about what was achievable outside the closed, proprietary model ecosystem. The implicit assumption in much of the coverage has been that American frontier AI labs — focused on subscription revenue and API monetization — would cede the open-weight space to international competitors willing to undercut on cost and licensing.</p>\n\n<p>Poolside's Laguna XS.2 release complicates that story. This is an American startup, backed by Bain Capital Ventures, Nvidia, and eBay Ventures, choosing to release a high-performing model under one of the most permissive open-source licenses available. Poolside's own blog framed the decision in direct terms, stating that <strong>"the West needs strong open-weight models."</strong></p>\n\n<p>The company's background makes this release more significant than a typical startup model drop. Poolside was founded in 2023 by Jason Warner, former CTO at GitHub, and Eiso Kant, founder of AI-for-code company source{d}. Prior to this launch, Poolside had operated almost exclusively in private and government/public sector contexts, building AI coding tools for high-security enterprise environments. This is the first time the company has made its models broadly available to the public.</p>\n\n<p>The timing also coincides with what appears to be a significant inflection point in Poolside's funding trajectory. The company raised a $500 million Series B round in October 2024, led by Bain Capital Ventures, with Nvidia and eBay Ventures participating — a round that valued Poolside at $3 billion at the time. More recently, Nvidia was reportedly preparing to invest between $500 million and $1 billion in Poolside as part of a $2 billion fundraising round that would value the company at $12 billion. That kind of capital, combined with a reported 10,000-GPU training cluster, gives Poolside resources that most open-model efforts cannot match.</p>\n\n<p>The Laguna models are also the product of a focused team: approximately 60 people in Poolside's Applied Research organization, spanning architecture, data, pre-training, and reinforcement learning. For a team of that size to produce two models trained on more than 30 trillion tokens — with benchmark results that outpace some larger competitors — is a meaningful signal about the efficiency of Poolside's research operation.</p>\n\n<h2>What Investors and Founders Are Saying</h2>\n\n<p>Poolside co-founder Jason Warner has previously framed the company's mission in expansive terms. <strong>"We believe software development will be the first broad capability where AI will reach and surpass human-level intelligence,"</strong> Warner has stated.</p>\n\n<p>Enrique Salem, a partner at Bain Capital Ventures, offered a view on the broader stakes of Poolside's work: <strong>"Their success will benefit everyone…advancing AI for software advances all of AI."</strong></p>\n\n<h2>What Comes Next for Poolside and Laguna</h2>\n\n<p>Poolside has not announced a specific roadmap for what follows the Laguna launch, but several near-term developments are already in motion. Both pool and shimmer — the terminal agent and cloud dev environment — are currently in preview, suggesting active user feedback will shape their development. The two products represent Poolside's attempt to build a complete agentic coding workflow around its models, rather than simply shipping weights and stepping back.</p>\n\n<p>The Apache 2.0 license on Laguna XS.2 means the model can be used commercially, fine-tuned, and redistributed without restriction — a decision that will likely accelerate community adoption and derivative work. How the broader developer ecosystem responds to XS.2's benchmark performance, and whether real-world agentic coding results match the benchmark numbers, will be a key test over the coming weeks.</p>\n\n<p>The free API access period for both Laguna models is described as time-limited, though Poolside has not specified an end date. Developers and teams looking to evaluate the models in live coding environments have a window to do so at no cost via the Poolside API or OpenRouter before any pricing changes take effect.</p>\n\n<p>Whether the larger Laguna M.1 — with its 225B total parameters and 23B active parameters — will receive similar open treatment remains to be seen. For now, the open-weight, consumer-hardware-friendly Laguna XS.2 is the model positioned for broad adoption, and its launch marks a clear signal that Poolside is ready to compete in the public market, not just the enterprise one.</p>\n\n<p>For more tech news, visit our <a href=\"/news\">news section</a>.</p>\n\n<h2>Why This Matters for Your Productivity</h2>\n\n<p>AI coding models like Laguna XS.2 are not just tools for professional developers — they represent a broader shift in how knowledge workers interact with software, automate repetitive tasks, and reclaim time for higher-order thinking. As capable agentic coding models become free and locally runnable, the productivity floor for technical and semi-technical professionals rises significantly. At Moccet, we track the tools and technologies that meaningfully shift what individuals can accomplish in their working hours. <a href=\"/#waitlist\">Join the Moccet waitlist to stay ahead of the curve.</a></p>", "excerpt": "American AI startup Poolside made its first-ever public model release on April 28, 2026, launching Laguna XS.2 — a free, open-weight agentic coding model with just 3 billion active parameters that outperforms Claude Haiku 4.5 and Gemma 4 31B on SWE-bench Pro. The model runs locally on a Mac with 36 GB of RAM and is available under an Apache 2.0 license on Hugging Face. The launch marks Poolside's debut in the public AI market after years of operating exclusively with government and enterprise clients.", "keywords": ["Poolside Laguna XS.2", "open-weight AI coding model", "agentic coding AI", "SWE-bench Pro", "open source AI model 2026"], "slug": "poolside-laguna-xs2-free-open-weight-agentic-coding-model" } ```