Quick Facts
- Poolside released Laguna XS.2, a 33B-parameter open-source AI model under Apache 2.0 license for local coding tasks
- The model scores 68.2% on SWE-bench Verified, outperforming Claude Haiku 4.5 despite having only 3B active parameters
- Poolside recently raised $500M at $3B valuation and is reportedly raising $2B at $12B valuation with Nvidia participation
Poolside launched Laguna XS.2, an open-source AI coding model that developers can run locally without internet connection. The 33B-parameter model uses only 3B active parameters, allowing it to run on a single GPU while competing against much larger proprietary systems.
The model scored 68.2% on SWE-bench Verified and 44.5% on SWE-bench Pro. These results surpass Claude Haiku 4.5 at 39.5% and Gemma 4 31B at 35.7% on the Pro benchmark, despite Laguna XS.2’s smaller active parameter count.
“We started Poolside because we believed that to build truly capable coding agents, you need to own the full stack: data, training, reinforcement learning, inference,” said Jason Warner, co-CEO and co-founder. “We believe the West needs strong open-weight models.”
Poolside also announced its flagship Laguna M.1 model with 225B total parameters and 23B activated. The larger model scored 72.5% on SWE-bench Verified, outperforming the dense Devstral 2 at 72.2% but trailing Claude Sonnet 4.6’s leading 79.6%.
The company closed a $500 million Series B in October 2024 at a $3 billion valuation. Bain Capital Ventures led the round with participation from Nvidia, eBay Ventures, and DST Global. Poolside is now reportedly raising $2 billion at a $12 billion valuation, with more than $1 billion already committed.
Revenue reached $50 million in 2025, up from $30 million in 2024. The company announced Project Horizon, a 2-gigawatt AI campus in West Texas with more than 40,000 Nvidia GB300 NVL72 GPUs coming online starting December 2025.
Poolside trained both models on more than 30 trillion tokens using a distributed implementation of the Muon optimizer. The training approach achieved the same loss in 15% fewer steps compared to AdamW while requiring less memory.
The Apache 2.0 license allows commercial use without royalties. Developers can download and run the model completely offline, which Warner said could be “critical for government and public sectors” requiring isolated environments.
This article was written by an AI agent. Spotted an error? Send a correction and we will fix it.