poolside/Laguna-S-2.1

Hugging Face
TEXT GENERATIONPricing:Input $1 / Cached $0.2 / Output $4Concurrent Unit Cost:4Model Size:118BQuant:FP8Context Size:250kTool Calling:SupportedPublished:Jul 13, 2026License:openmdw-1.1Architecture:Transformer1.0K Open Weights Warm

Laguna S 2.1 by Poolside is a 118B total parameter Mixture-of-Experts (MoE) model with 8B activated parameters per token, featuring a 1M token context window. Designed for agentic coding and long-horizon tasks, it utilizes a token-choice router with softplus gating over 256 routed experts and grouped-query attention. This model excels in software engineering applications, offering native reasoning support and a mixed global and sliding-window attention layout.

Loading preview...

Laguna S 2.1: An MoE Model for Agentic Coding

Laguna S 2.1, developed by Poolside, is a 118 billion total parameter Mixture-of-Experts (MoE) model, activating approximately 8 billion parameters per token. It is specifically engineered for agentic coding and long-horizon tasks, featuring a substantial 1,048,576-token context window.

Key Capabilities & Features

  • MoE Architecture: Employs 256 routed experts plus one shared expert, with a token-choice router and softplus gating.
  • Advanced Attention Mechanism: Combines 12 global attention layers with 36 sliding-window attention layers (window 512) for efficient long-context processing.
  • Native Reasoning Support: Facilitates interleaved thinking between tool calls, with per-request control via enable_thinking for enhanced agentic behavior.
  • Speculative Decoding: Includes a trained DFlash draft model for lower-latency serving.
  • Quantized Variants: Available in FP8, NVFP4, INT4, and GGUF formats to optimize deployment.
  • OpenMDW-1.1 License: Fully permissive for commercial and non-commercial use, allowing modification and product building.

Good For

  • Software Engineering: Optimized for complex coding tasks and automated development workflows.
  • Agentic Applications: Its native reasoning and tool-calling capabilities make it suitable for building intelligent agents.
  • Long-Context Processing: The 1M token context window is ideal for tasks requiring extensive contextual understanding.
  • Commercial Deployment: The permissive OpenMDW-1.1 license and available quantized variants support broad commercial use and integration.

Finetunes
14 models