Poolside Laguna S 2.1: 118B Parameters on Locally on DGX Spark

In this video, I test Poolside’s Laguna S 2.1 locally on a single NVIDIA DGX Spark. Laguna S 2.1 is a 118B-parameter Mixture-of-Experts model with roughly 8B parameters activated per token. I run the native NVFP4 checkpoint using vLLM with DFlash speculative decoding, then put the model through practical BenchLocal tests. I cover: Model download and setup Loading Laguna S 2.1 with vLLM DGX Spark compatibility Startup time and performance Instruction-following tests Coding and reasoning capability Local AI advantages and limitations Whether this 118B model is actually practical on one machine This is a fully local test—no cloud inference and no external API generating the model responses. System used: NVIDIA DGX Spark 128 GB unified memory Laguna S 2.1 NVFP4 vLLM 0.25.1 DFlash speculative decoding BenchLocal testing interface Poolside publishes Laguna S 2.1 as an open-weight model under the OpenMDW-1.1 license. Subscribe for more real-world tests of large open models running locally on the DGX Spark. #LagunaS21 #PoolsideAI #DGXSpark #LocalAI #OpenSourceAI #OpenWeights #vLLM #AIModels #CodingAI #NVIDIA #LLM #ArtificialIntelligence #BenchLocal #TechYouTube