Poolside's Laguna S 2.1 is the West's most capable open-weight coder
118B params that fit on one DGX Spark, top SWE-bench for their class, and run free to 256K on OpenRouter — plus Google's new vuln-hunting Gemini.
Copy markdown
Run a frontier-class coder on one box
Poolside's Laguna S 2.1 is a 118B open-weight coding model with a 1M-token context, released on Hugging Face in BF16, FP8, INT4 and NVFP4 builds. The 4-bit weights need only ~59GB, so the whole model fits on a single NVIDIA DGX Spark — a self-hostable frontier-class coder that used to demand a cluster.
It punches 10x above its weight
Laguna S 2.1 scores 70.2% on Terminal-Bench 2.1, 59.4% on SWE-Bench Pro (public) and 78.5% on SWE-Bench Multilingual, matching or beating models several times its size from DeepSeek and Nvidia. It even independently proved Erdos Problem #397 — and Poolside trained it on 4,000 H200s in under four weeks.
Free to 256K, cents past that
On OpenRouter it's free up to 256K context, then $0.10 / $0.20 / $0.01 per 1M input / output / cache-read tokens at the full 1M — and it's also live on Baseten, Kilo, Prime Intellect and ZML. The permissive OpenMDW-1.1 license lets you self-host and ship it commercially, no usage gate.
Gemini 3.5 Flash Cyber hunts your bugs
Google's new lightweight security model finds, validates and patches vulnerabilities: it surfaced 55 unique issues in the V8 JavaScript engine versus Claude Opus 4.6's 36, and a Google team used it to find production RCEs in two hours. The catch — it's gated to governments and trusted partners via CodeMender for now, with the capability rolling into the Gemini Enterprise Agent Platform.