Poolside Releases Laguna S 2.1 — an Open-Weight Model for Agentic Coding That Punches Above Its Weight Class in SWE-Bench Multilingual

Poolside has released Laguna S 2.1, an open-weight model for agentic coding with 118 billion parameters, of which about 8 billion are activated per token thanks to a Mixture-of-Experts (MoE) architecture. The model supports a context of up to 1 million tokens in both reasoning and non-reasoning modes. Weights are available on Hugging Face under the OpenMDW-1.1 license, and the model size allows it to run on a single NVIDIA DGX Spark. On long-context coding benchmarks, Laguna S 2.1 achieves results comparable to models several times larger, including DeepSeek-V4-Pro-Max, Nemotron 3 Ultra, and Inkling. For example, on SWE-Bench Multilingual, the model scored 78.5%, which is the best published result among open-weight models. On Terminal-Bench 2.1 with reasoning mode enabled, the result was 70.2%. Meanwhile, on the DeepSWE v1.1 test, the model scored 40.4% versus 9.0% for DeepSeek-V4-Pro-Max, with approximately six times fewer active parameters. The model has two reasoning modes: off and maximum (default). In maximum mode, the model determines the compute budget autonomously, yielding an improvement on DeepSWE from 16.5% to 40.4%, but at the cost of increased token count (from 99k to 249k). Poolside's team has published three unedited trajectories of the model's work: creating a working HTML/CSS browser engine from an empty folder, optimizing its own agentic engine (5.2% speedup and 71% memory allocation reduction), and independently rediscovering a solution to Erdős problem #397 in Perl. For deployment, all 118 billion parameters remain in memory. With 4-bit quantization (NVFP4 or INT4), the weights occupy about 59 GB, fitting into the 128 GB unified memory of a single DGX Spark. With FP8, about 118 GB (one Spark or H200), and with BF16, 236 GB (two Sparks or a multi-GPU node). Training began on May 22, 2026, on 4096 NVIDIA H200 GPUs and took less than nine weeks. The model is also available through OpenRouter, vLLM, SGLang, Ollama, and other platforms.