Laguna S 2.1 is live on Kilo

Poolside just shipped Laguna S 2.1, and you can run it in Kilo today across the IDE extension, CLI, and Cloud Agents. Try it for free for a limited time!Poolside builds open-weight foundation models and they share Kilo’s commitment to building in the open. Their previous models have been super popular with developers at the cutting edge, and we’re excited for our community to try the new Laguna S model.

Here’s why Laguna S 2.1 is worth your attention: it’s a 118B parameter model that only activates 8B per token, and it’s already going head to head with models many times its size. On SWE-bench Multilingual it hits 78.5%, which puts it right in the mix with Tencent Hy3 (295B-A21B) and DeepSeek-V4-Pro Max (1.6T-A49B). On Toolathlon Verified it scores 49.7%, ahead of Nemotron 3 Ultra and DeepSeek-V4-Flash Max. That’s a lot of capability packed into a model small enough to run locally on a single NVIDIA DGX Spark.

The thing that actually matters: it doesn’t quit

Benchmarks are one thing, but the more interesting story with Laguna S 2.1 is behavioral. Poolside reports it holding onto a task, using tools, checking its own work, and recovering from failed approaches for up to 24 hours with little to no intervention. That’s the kind of long-horizon persistence that separates a model you babysit from one you actually delegate to.

A few examples from Poolside’s internal testing give a sense of what that looks like in practice:

It built a full HTML and CSS rendering engine over 181 reasoning turns, comparing its own output against a real browser pixel by pixel until it landed within three pixels of the reference. In a restricted sandbox with no internet access, it independently rediscovered a known solution to Erdős Problem 397, and when Python wasn’t available, it switched to Perl and derived a distinct parametric family on its own. It also spent an overnight run improving on a circle-packing result from DeepMind’s AlphaEvolve, and kept nudging the number closer to opt…