Laguna S 2.1 is a new 118B parameter Mixture-of-Experts (MoE) model purpose-built for agentic software engineering and long-horizon tasks. Sitting comfortably between the XS and M.1 models in the Laguna family, it achieves high efficiency by activating just 8B parameters per token.
More info is available on AImodels.fyi here!
Highlights
-
Massive 1M context window: Can process up to 1,048,576 tokens, making it ideal for ingesting large codebases and managing extended debugging sessions.
-
Native reasoning engine: Features “interleaved thinking,” allowing the model to reason before and between tool calls. (Tip: It performs best when prior “thinking” blocks are preserved in the chat history).
-
Unique architecture: Utilizes a token-choice router with 256 routed experts (plus one shared expert) and a 1:3 mix of global to sliding-window attention layers.
-
Strong benchmark performance: Puts up highly competitive numbers on coding-heavy evaluations like SWE-bench Multilingual (78.5%) and Terminal-Bench 2.1 (70.2%), holding its own against models with significantly larger active parameter counts.
-
Broad deployment support: Works out-of-the-box with popular serving engines including vLLM, SGLang, TRT-LLM, llama.cpp, and Ollama.
-
Fast & accessible: Includes a trained DFlash draft model for low-latency speculative decoding, plus multiple quantized variants (FP8, NVFP4, INT4, GGUF) to drastically reduce the ~236GB VRAM requirement of the base model.
License and availability:
Released under the permissive OpenMDW-1.1 license, Laguna S 2.1 is free to use and modify for both commercial and non-commercial applications.
