Poolside Releases 118B Open Coding Model Laguna S 2.1

Poolside released Laguna S 2.1 on May 22, 2026, an 118 billion-parameter open-weight model designed for agentic coding tasks. This Mixture-of-Experts (MoE) model activates approximately 8 billion parameters per token, allowing it to perform comparably to models many times its size on long-horizon coding benchmarks. The model supports a context window of up to 1 million tokens and is available on Hugging Face under an OpenMDW-1.1 license. Its architecture allows it to run on a single NVIDIA DGX Spark, making it accessible for deployment.

Laguna S 2.1 demonstrates strong performance, achieving a score of 70.2% on the Terminal-Bench 2.1 with thinking enabled, placing it at the top among open, disclosed-size models on Poolside's leaderboard. It also achieved 78.5% on SWE-Bench Multilingual, outperforming all other published models. The model is a scaled-up version of the Laguna XS family, trained on the same pre-training data as XS 2.1. Poolside has made the weights available in various precisions, including BF16, FP8, INT4, and NVFP4, along with official GGUF and MLX conversions.

The development of Laguna S 2.1 was rapid, with the entire process from the start of training to launch completed in under nine weeks. Pre-training commenced on May 22, 2026, utilizing 4,096 NVIDIA H200 GPUs. Notably, this is the first Poolside model where reinforcement learning was conducted using FP8 precision. The model's efficiency stems from its sparse activation, where only a fraction of its total parameters are utilized per token, while all 118 billion parameters remain in memory. This approach enables a mid-size model to exhibit the capabilities of a larger one while maintaining cost-effective serving.