Release2026-08-03
Poolside released Laguna S 2.1, an openly licensed coding model of 118 billion parameters that activates about 8 billion of them per request. It can hold roughly one million tokens of text in view at once and is aimed at software work that runs over many steps with tool use. The full-precision version needs several GPUs, though smaller compressed versions and hosted access are available.
What changed
Poolside's open Laguna line offered a smaller 33B model and a much larger 225B one, with nothing in between.
What it unlocks
Running a coding-focused model with a one-million-token context on self-managed hardware, or through hosted gateways, under a permissive commercial licence.
- 118B total, 8B active per token
- 1,048,576-token context
- 70.2% Terminal-Bench 2.1
- 78.5% SWE-bench Multilingual
What you need to act on it
- multiple GPUs (~236GB for full-precision weights) or a quantized variant
- or access via OpenRouter, Vercel AI Gateway or Ollama
Sources