llama.cpp Adds Support for Poolside's Laguna XS.2 and M.1 MoE Models
A new pull request in the llama.cpp repository introduces support for Poolside's Laguna XS.2 and M.1 Mixture-of-Experts (MoE) models. These models are specifically optimized for local agentic coding and long-horizon software engineering tasks. This integration allows developers to run Poolside's advanced coding models locally with optimized inference, bringing powerful agentic capabilities to consumer hardware. It expands the ecosystem of open-weights models available for local software development workflows. Laguna XS.2 is a 33B parameter model (3B active) utilizing Sliding Window Attention with per-head gating in 30 of its 40 layers to minimize KV cache requirements. Meanwhile, Laguna M.1 is a massive 225B parameter model (23B active) featuring a 256K context window and support for tool calling.
## BACKGROUND
llama.cpp is a popular open-source framework designed for efficient LLM inference on local hardware using C/C++. Mixture-of-Experts (MoE) architectures activate only a fraction of their total parameters per token, enabling high performance without requiring massive compute for every step. Sliding Window Attention further optimizes memory by limiting how far back tokens look in the context history, reducing the KV cache size.