poolside/Laguna-S-2.1
Laguna S 2.1 by Poolside is a 118 billion total parameter Mixture-of-Experts (MoE) model, activating 8 billion parameters per token, with a 1,048,576-token context window. It features a unique mixed global and sliding-window attention layout and native reasoning support. This model is specifically designed and optimized for agentic coding and long-horizon software engineering tasks.
Loading preview...
Overview
Laguna S 2.1 is a 118 billion total parameter Mixture-of-Experts (MoE) model developed by Poolside, utilizing 8 billion activated parameters per token. It is part of the Laguna series, featuring a token-choice router with softplus gating over 256 routed experts and one shared expert. The model incorporates grouped-query attention and an interleaved full/sliding-window attention mechanism, with a substantial 1,048,576-token context window.
Key Capabilities
- Extended Context Window: Supports a 1 million token context window for handling extensive codebases and complex tasks.
- Native Reasoning: Includes built-in support for interleaved thinking between tool calls, with per-request control via
enable_thinking. - Mixed Attention Layout: Employs a unique architecture with 12 global attention layers and 36 sliding-window attention layers (window 512).
- Speculative Decoding: Offers a trained DFlash draft model for reduced latency serving.
- Quantized Variants: Available in FP8, NVFP4, INT4, and GGUF formats for flexible deployment.
Benchmarks
Laguna S 2.1 demonstrates strong performance in agentic coding benchmarks, achieving 70.2% on Terminal-Bench 2.1, 78.5% on SWE-bench Multilingual, and 59.4% on SWE-Bench Pro (Public Dataset), often outperforming larger models in its class.
Good For
- Agentic Coding: Specifically designed for automated software engineering tasks.
- Long-Horizon Work: Its large context window and reasoning capabilities make it suitable for complex, multi-step projects.
- Tool Use: Optimized for scenarios requiring tool integration and sequential decision-making.