Release2026-07-21
Poolside released Laguna S 2.1, an openly downloadable coding model that is far smaller than most rivals yet scores close to them on tests of long, multi-step programming work. The company published every evaluation run behind its published scores. It also lists known faults, including trouble following tool definitions in other companies' agent software and a tendency to think for too long before acting.
What changed
Poolside's earlier Laguna models were smaller and, in the company's own account, more prone to stopping early on partly finished work.
What it unlocks
Running a coding assistant that works through long multi-step tasks on a single workstation-class machine rather than calling a large hosted model.
- 118 billion total parameters with 8 billion active per token
- context window of up to 1M tokens
- Terminal-Bench 2.1: 70.2% with thinking, 60.4% without
- DeepSWE v1.1: 40.4% with thinking, 16.5% without
- trained from scratch to launch in under nine weeks on 4,096 NVIDIA H200 GPUs
What you need to act on it
- hardware able to run a 118B-parameter open-weights model, or use of Poolside's hosted chat
- downloading the weights from Hugging Face for self-hosting
Sources