Laguna XS.2 and M.1
poolside.ai
April 28, 2026
16 min read
🔥🔥🔥🔥🔥
50/100
Summary
Laguna M.1 and Laguna XS.2 are the first two models in the Laguna family, released alongside a runtime for training and operating agents. Open weights will be available soon, with collaboration from NVIDIA for model building, data, automixing, and agent reinforcement learning.
Key Takeaways
- Laguna M.1 is a 225 billion parameter Mixture of Experts model that completed pre-training at the end of last year, achieving 46.9% on SWE-bench Pro and 40.7% on Terminal-Bench 2.0.
- Laguna XS.2 is the first open-weight model in the Laguna family, featuring 33 billion total parameters and reaching 44.5% on SWE-bench Pro and 30.1% on Terminal-Bench 2.0, with weights available under an Apache 2.0 license.
- Both Laguna M.1 and Laguna XS.2 are designed for agentic coding tasks and are available for free use through an API and OpenRouter during a limited preview period.
- The models were developed by a team of approximately 60 researchers focusing on architecture, data, pre-training, and reinforcement learning.
Coverage & discussions
Hacker News · poolside.ai
Community discussions
Community Sentiment
Positives
- The emergence of Laguna XS.2 and M.1 showcases competitive innovation in the AI space, indicating a healthy evolution of local LLMs.
- Quantization techniques are making smaller models increasingly viable on consumer hardware, which could democratize access to advanced AI capabilities.
- Despite some concerns, the ongoing development in local LLMs suggests exciting potential for future advancements and applications.
Concerns
- Laguna XS.2 and M.1 are perceived as underperforming compared to Qwen3.6, raising questions about their competitiveness in the current landscape.
- The disparity in benchmark scores between Laguna models and Qwen3.6 indicates that they may not meet the expectations for performance in their class.
- There are concerns about the fragmentation of AI research efforts among smaller companies, which could dilute resources and hinder progress.
![[AINews] Why OpenAI Should Build Slack](https://substackcdn.com/image/fetch/$s_!XQAE!,w_1200,h_675,c_fill,f_jpg,q_auto:good,fl_progressive:steep,g_auto/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F89ee056a-0ea2-4473-8e1c-9b21f034c717_1474x2116.png)
