Laguna M.1 is a large sparse Mixture-of-Experts transformer designed for agentic coding and long-horizon software tasks. Its architecture pairs a 225-billion-total-parameter footprint with 23 billion activated parameters per token, giving it broad capacity while keeping per-token compute manageable. The model follows a 70-layer layout that begins with three dense SwiGLU layers and then transitions into 67 sparse MoE layers using 256 experts with top-16 routing and auxiliary-loss-free load balancing. Attention runs globally across every layer with 64 query heads and 8 key-value heads, paired with softplus attention output gating, an arrangement that favors consistent context handling across long agentic traces rather than localized sliding-window patterns.

The model is positioned for software engineering agents that interleave reasoning with tool use, and it is accompanied by a release blog and a technical report detailing Poolside's Model Factory training approach, post-training recipe, and asynchronous off-policy reinforcement learning from code agents. Poolside describes the broader Laguna family as a tiered set of open-weight coding models, with Laguna M.1 sitting as the high-capacity variant alongside smaller siblings aimed at lighter or on-device deployment. The combination of a sparse MoE backbone, native support for interleaved thinking between tool calls, and open-weight availability under an Apache 2.0 license makes Laguna M.1 well suited to teams building self-directed coding agents that need to sustain long, tool-heavy workflows.