OpenRouter Hosts Poolside AI Laguna Models for Free Agentic Coding

Apr 28, 2026 · Updated May 9, 2026

OpenRouter added the first public models from Poolside AI, a frontier lab building foundation models for software agents, to its unified API. The release includes Laguna M.1, a flagship model for complex engineering, and Laguna XS.2, a 33B Mixture of Experts (an architecture that activates only a subset of parameters per token) model.

Context window: 128K tokens
Output limit: 8K tokens
Architecture (XS.2): 33B Mixture of Experts
Active parameters (XS.2): 3B tokens
Precision: fp8 quantized
Pricing: Free for a limited time

While most developers use general-purpose models for coding, these models are built from scratch for the agentic loop. They prioritize tool calling and reasoning over chat fluency to handle multi-step tasks. This release follows a surge in specialized reasoning models and long-context agentic tools appearing on the platform.

Access both models via the OpenRouter API or download the weights from Hugging Face. Both feature a 128K context window and are quantized to fp8 (a 8-bit floating point format for faster inference) to reduce latency. Access on OpenRouter is currently free for a limited time.

View the full update on openrouter.ai

OpenRouter

@OpenRouterApr 28

The first public foundation models from @poolsideai just dropped on OpenRouter! Laguna M.1 and Laguna XS.2. Built from scratch for agentic coding and long-horizon work. Free for a limited time ⬇️ https://t.co/toDCzLsbgu

20249

View on X

Still wondering? A few quick answers below.

Laguna M.1 and Laguna XS.2 are the first public foundation models from Poolside AI, a lab focused on software engineering agents. Unlike general-purpose models, these are built from scratch for agentic coding and long-horizon tasks. They prioritize tool calling and reasoning capabilities to help autonomous agents navigate complex codebases and execute multi-step software development workflows.

Laguna XS.2 uses a 33B parameter Mixture of Experts architecture, where only 3B parameters are active during any single inference turn. This design allows the model to maintain high reasoning performance for coding tasks while remaining computationally efficient. It is quantized to fp8 precision, which further optimizes it for fast and cost-effective execution in agentic loops.

While not strictly open source in terms of training data, Poolside has released the model weights for both Laguna M.1 and Laguna XS.2 on Hugging Face. This open-weight release allows developers to download and run the models on their own infrastructure. Additionally, the models are available for public use via the OpenRouter API for integrated workflows.

Both Laguna models feature a 128,000 token context window, allowing them to process large segments of code or multiple files in a single request. They also support an output limit of up to 8,000 tokens. This large capacity is specifically designed for long-horizon work where an agent needs to maintain state across complex engineering tasks.

You can access Laguna M.1 and Laguna XS.2 through the OpenRouter API, where they are currently available for free for a limited time. Developers can also find the model weights on Hugging Face for local deployment. The models are optimized for integration into coding agents, IDEs, and CI/CD workflows that require specialized reasoning for software engineering.