en.Wedoany.com Reported - Recently, the US AI startup Poolside (founded in San Francisco in 2023) released its Laguna series of models, including the proprietary model Laguna M.1 and the open-source model Laguna XS.2. The Laguna XS.2 is released under the Apache 2.0 license, has a total of 33 billion parameters (3 billion active), and is specifically designed for local intelligent programming, allowing developers to run it on a single GPU without an internet connection.

Poolside simultaneously launched two tools: a terminal-based coding agent called "pool" and a cloud IDE called "shimmer," the latter of which supports programming on mobile devices. Post-training engineer George Grigorev stated, "We think we can deploy models to enterprise customers faster, and we can ship weights in a completely isolated, on-premise environment, enabling offline work."

Benchmarks show that Laguna XS.2 achieves a score of 44.5% on SWE-bench Pro, closely approaching M.1's 46.9%, and surpasses models like Claude Haiku 4.5. In terms of training, the model was trained from scratch using the Muon optimizer and the AutoMixer data filtering system on 30 trillion tokens, 13% of which were synthetic data.

Running Laguna XS.2 requires at least 36GB of unified memory (Apple Silicon) or 24-32GB of VRAM (PC, such as an RTX 5090). The model is available for download on Hugging Face and is suitable for consumer-grade hardware. Poolside emphasizes open-source to support the community and hopes to build an open AI ecosystem.

Poolside believes that software development is a key pathway to general intelligence. The team has focused intensely on software engineering through tens of thousands of experiments over three years, using reinforcement learning for self-improvement. This release provides developers with the open-source Laguna XS.2 model for efficient local programming.

This article is compiled by Wedoany. All AI citations must indicate the source as "Wedoany". If there is any infringement or other issues, please notify us promptly, and we will modify or delete it accordingly. Email: news@wedoany.com