Model Releases
Poolside Laguna-M.1: The 225B Coding Giant Arrives in 2026
Poolside unveils Laguna-M.1, a massive MoE model designed for agentic software engineering with 225B parameters.
April 28, 2026
Model ReleaseLaguna-M.1
Introduction: A Milestone in AI History
Poolside has officially unveiled Laguna-M.1, a monumental leap in the landscape of generative AI for software engineering. Released on April 28, 2026, this model represents a significant milestone for the company and the industry at large. It is not merely an incremental update but a foundational shift in how agentic coding is approached. For developers, this means a new standard for handling complex, long-horizon tasks that previous models struggled with. The model is part of the broader Laguna family, designed to push the boundaries of what large language models can achieve in real-world production environments.
Its architecture is built on lessons learned from previous iterations, specifically focusing on efficiency and raw capability. As the most capable model to date, it completes pre-training at the end of 2025. The release signifies a transition from experimental agentic tools to enterprise-grade solutions. The Laguna family is now anchored by this powerhouse, setting the stage for future releases that will likely inherit this robust foundation.
- Released on 2026-04-28 by Poolside.
- Most capable model to date in the Laguna family.
- Foundation for the entire Laguna model family.
- Completed pre-training at the end of 2025.
Key Features & Architecture
The architecture of Laguna-M.1 is a testament to high-scale training. It operates as a 225B parameter Mixture-of-Experts model, but it only activates 23B parameters per token. This MoE design allows for massive capacity without prohibitive inference costs. The model was trained from scratch on 30T tokens using the Muon optimizer. It utilized 6,144 interconnected NVIDIA Hopper GPUs entirely in-house. The context window supports 128K inputs with up to 8K output tokens. This setup enables the model to handle massive codebases and extensive documentation contexts.
Beyond standard inference, the model incorporates a custom async on-policy RL system with an Agent Client Protocol (ACP) server. This infrastructure supports advanced reasoning capabilities required for complex software tasks. The 128K context window is particularly vital for developers managing large legacy codebases where information density is high.
- 225B total parameters with 23B activated per token.
- 128K context window with 8K output tokens.
- Trained on 30T tokens using Muon optimizer.
- Custom async on-policy RL system with ACP server.
Performance & Benchmarks
Performance metrics are crucial for any serious engineering model. On SWE-bench Verified, Laguna-M.1 achieves a score of 72.5%. It scores 67.3% on SWE-bench Multilingual and 46.9% on SWE-bench Pro. Terminal-Bench 2.0 sees a score of 40.7%. These numbers place it above many competitors in the current landscape. The model is specifically tuned for agentic workflows, meaning it can plan and execute multi-step tasks autonomously.