Daily AI CatchupPoolside's Laguna S 2.1 Leads Western Open-Source Coding
Poolside launched Laguna S 2.1, a 118B Mixture-of-Experts coding model that activates only 8B parameters per token — small enough to run on a single Nvidia DGX Spark desktop, with weights free on Hugging Face. It tops Western open-source coding benchmarks, though it still trails China's Kimi K3 significantly. Training took under nine weeks.
Why it matters
💻 Developer · Free weights on Hugging Face and desktop-GPU-friendly sizing make this worth trying locally if you want an open coding model without renting a cluster — just temper expectations against Kimi K3.
📦 Product · An open-source coding model good enough to self-host changes the build-vs-buy math for any internal dev-tooling feature that currently depends on a hosted API.
🎨 Design · Not directly design-relevant, but a reminder that "good enough locally-run" models are closing in on cloud-hosted quality for specialized tasks.
📈 Business · Nine weeks of training to reach the top of Western open-source coding is a fast turnaround — a data point on how quickly a well-funded challenger can catch up on a narrow benchmark.
🤔 Just Curious · A startup called Poolside released a free AI coding model small enough to run on a single desktop computer, and it's now the best of its kind outside of China.
Try this: If you have a DGX Spark or equivalent desktop GPU box, pull Laguna S 2.1 from Hugging Face and run it against your usual coding benchmark set to see where it lands versus your current model.
Sources: Poolside: Introducing Laguna S 2.1