Poolside's Tiny Laguna S 2.1 Beats a Giant Rival on Coding Benchmarks by a 4x Margin

Poolside just shipped a coding model with 8 billion active parameters that outscores a rival model roughly six times its computational size. Laguna S 2.1, the startup's newest open-weight release, hits 40.4% on the DeepSWE v1.1 benchmark, dwarfing DeepSeek-V4-Pro-Max's 9.0% despite activating a fraction of the parameters per token. It's the kind of efficiency gap that tends to get noticed in a market where inference cost, not just raw capability, decides which models actually get deployed.

The model is a mixture-of-experts system with 118 billion total parameters but only 8 billion active at inference time, small enough to run on a single NVIDIA DGX Spark. On SWE-Bench Multilingual, it posts 78.5%, topping every other model in Poolside's published comparison table. Terminal-Bench 2.1 results land at 70.2% with the model's "thinking mode" switched on, up from a 60.4% baseline without extended reasoning. That reasoning mode ships enabled by default, and it's doing real work: on DeepSWE specifically, turning it on pushes accuracy from 16.5% to 40.4%, though that jump comes at a cost. Trajectories that use thinking mode run roughly 249,000 completion tokens, versus about 99,000 tokens without it, meaning the accuracy gain is bought with roughly 2.5 times the token spend.

Laguna S 2.1 supports context windows up to 1 million tokens in both thinking and non-thinking configurations, and it's Poolside's first model where the reinforcement learning stage ran in FP8 precision, a choice that likely contributed to the compressed training timeline. Pre-training started on May 22, 2026, using 4,096 NVIDIA H200 GPUs, and the entire training-to-launch cycle wrapped in under nine weeks. The model is built as a scale-up of the Laguna XS family, sharing its pre-training data with XS 2.1, and weights are published on Hugging Face under Poolside's OpenMDW-1.1 license.

From Government Contracts to Open Weights

The release marks a notable shift for Poolside, a company that built its early business around government and public-sector clients before pivoting toward open releases, starting with XS.2 under an Apache 2.0 license. Laguna S 2.1 is now the third coding model the company has put out in roughly three months, following Laguna M.1 and XS.2 in April 2026, an unusually fast cadence for a startup competing against labs with far deeper compute budgets.

Poolside's own demonstrations lean into the model's ability to handle open-ended, long-horizon tasks. In one showcase, Laguna S 2.1 built a functioning browser engine from an empty folder in 50 minutes, one capable of rendering HTML and CSS. In another, it produced a proof for Erdős Problem #397, a number theory question that had sat unresolved since 1975, completing the task in 40 reasoning steps for a compute cost of $0.088.

Those anecdotes make for good marketing, but the benchmark numbers are what should matter to engineering teams evaluating self-hosted coding assistants. A model that fits on a single DGX Spark while beating far larger competitors on SWE-Bench Multilingual changes the calculus for teams that want strong code generation without renting a fleet of GPUs.