Laguna M.1

Laguna M.1 is a large sparse mixture-of-experts transformer built specifically for agentic coding and long-running software workflows. Its 70-layer architecture pairs three dense SwiGLU layers at the input with 67 sparse MoE layers using 256 experts, top-16 routing, and an auxiliary-loss-free load-balancing scheme, yielding 225B total parameters while activating only 23B per token. The model uses global attention throughout, with 64 query heads and 8 key-value heads combined with softplus attention output gating, a design choice aimed at keeping context coherent across the kinds of multi-step, tool-driven sessions that coding agents run.

The release emphasizes practical agentic behavior: interleaved thinking that can be toggled between or within tool calls, and competitive results against state-of-the-art open-weight and frontier models on SWE-bench Verified, SWE-bench Multilingual, SWE-Bench Pro, and Terminal-Bench 2.0. Because the weights are published openly under an Apache 2.0 license alongside a detailed technical report and post-training recipe, teams can self-host and adapt Laguna M.1 for internal codebases, long refactors, and retrieval-heavy debugging pipelines where sustained multi-turn reasoning matters more than raw single-prompt fluency.