Description
Laguna S 2.1 is a sophisticated Mixture-of-Experts model developed for agentic coding and long-horizon work. With a total of 118 billion parameters and 8 billion activated parameters per token, it is part of the Laguna series, positioned between Laguna XS 2.1 and Laguna M.1. The model employs a token-choice router with softplus gating over 256 routed experts and one shared expert, along with grouped-query attention and interleaved full/sliding-window attention.
Read More
Released under the OpenMDW-1.1 license, Laguna S 2.1 allows users to freely use, modify, and build commercial products without requiring permission. It is optimized for software engineering tasks, offering native reasoning support and speculative decoding through a trained DFlash draft model. The model supports quantized variants such as FP8, NVFP4, INT4, and GGUF, reducing the need for extensive hardware resources.
Laguna S 2.1 features a mixed SWA and global attention layout with 48 layers in a 1:3 global-to-SWA ratio. It provides a 1,048,576-token context window, enabling extensive context management. The model's vocabulary consists of 100,352 tokens, utilizing the Laguna family tokenizer for text-to-text modality.
Benchmark results demonstrate Laguna S 2.1's competitive performance across various evaluations, including Terminal-Bench 2.1 and SWE-bench Multilingual. It integrates seamlessly with engines like vLLM, SGLang, TRT-LLM, and llama.cpp, offering flexibility in deployment and usage.
Designed for software engineering and agentic coding use cases, Laguna S 2.1 is subject to Poolside's Acceptable Use Policy, ensuring responsible application. Users are advised to maintain safety guardrails and report any security vulnerabilities to Poolside.
- CategoryAI Models & LLMs
- Pricing ModelOpen source
- DeploymentCloud
- PlatformsWeb
Laguna S 2.1 Highlights
118B total parameters
8B activated parameters per token
Mixture-of-Experts model
OpenMDW-1.1 license
1,048,576-token context window
Speculative decoding with DFlash
Quantized variants available
Native reasoning support
Getting Started with Laguna S 2.1
Access model page: Visit the Hugging Face website
Authenticate: Log in to your account
Configure: Set up model parameters and options
Send prompts: Input text for processing
Fine-tune: Adjust model for specific tasks
Laguna S 2.1's Use Cases
- Agentic Coding
- Long-Horizon Tasks
- Software Engineering
- Commercial Product Development
- Speculative Decoding
