Laguna-M.1: 225B MoE Model Released

Laguna-M.1, developed by Poolside, is a Mixture-of-Experts (MoE) model with a total of 225B parameters, using 23B active parameters per token. It is designed specifically for agentic coding and long-horizon work.

Key technical features are as follows:

• MoE structure: 67 out of 70 layers are sparse MoE layers, supporting 256 Experts with top-k=16 routing.
• Architecture: Uses a global attention architecture, applying softplus attention output gating.
• Inference support: Supports Interleaved thinking between tool calls, allowing reasoning to be enabled or disabled on request.
• Context window: Supports a long context of 262,144 tokens.

In terms of performance, Laguna-M.1 shows strong results competitive with the latest open-weight models on major agentic benchmarks, including SWE-bench Verified (74.6%) and SWE-bench Multilingual (63.1%). The model is distributed under the Apache 2.0 license, allowing commercial use.