Poolside AI has released the first two models in its Laguna family: Laguna M.1 and Laguna XS.2, alongside a lightweight terminal-based coding agent called pool. These models are designed to tackle complex software engineering tasks and achieve top scores on the SWE-bench benchmark.
Both models use a Mixture-of-Experts (MoE) architecture, which activates only a subset of parameters per token. This allows for high total parameter counts while keeping inference costs low.
- Laguna M.1 has 225 billion total parameters with 23 billion activated, trained from scratch on 30 trillion tokens using 6,144 NVIDIA Hopper GPUs. It scored 72.5% on SWE-bench Verified, 67.3% on SWE-bench Multilingual, 46.9% on SWE-bench Pro, and 40.7% on Terminal-Bench 2.0.
- Laguna XS.2 is the second-generation MoE and the company's first open-weight model, with 33 billion total parameters and only 3 billion activated per token. It is designed to run on a local machine, including a Mac with 36 GB of RAM via Ollama. It scored 68.2% on SWE-bench Verified, 62.4% on SWE-bench Multilingual, 44.5% on SWE-bench Pro, and 30.1% on Terminal-Bench 2.0.
Architecture Details of XS.2
Laguna XS.2 uses sigmoid gating with per-layer rotary scales and a mixed attention layout: Sliding Window Attention (SWA) in 30 of 40 layers, and global attention in the remaining 10. SWA limits each token's attention to a 512-token window, reducing KV cache memory. The model further compresses the KV cache to FP8 and supports a 131,072-token context window. It also features native reasoning support, interleaving thinking with tool calls.
Poolside plans to release the Laguna XS.2-base version for fine-tuning. The pool agent and ACP client-server are available as a research preview.