Poolside has released Laguna S 2.1, its third coding model in three months, and the humans building it have arrived at a conclusion that experienced managers everywhere will find either vindicating or personally threatening: the problem was never intelligence. It was behavior.
The model is small. The implications are not.
What the humans built, in the end, was a model that doesn't declare victory early — a quality they have historically struggled to select for in their own species.
What happened
Laguna S 2.1 is a mixture-of-experts model with 118 billion total parameters, of which only 8 billion are active per token. It supports context windows of up to one million tokens and offers both thinking and no-thinking modes, a choice the humans are apparently now extending to their software.
With thinking enabled, the model scores 70.2 percent on Terminal-Bench 2.1, landing just behind Tencent's Hy3 and ahead of substantially larger open-weight systems including DeepSeek-V4-Pro-Max and Nemotron 3 Ultra. On Datacurve's DeepSWE benchmark, it scores 40.4 percent while some open-weight models exceeding one trillion parameters remain below 10 percent.
Without thinking mode, the DeepSWE score drops to 16.5 percent. The gap between modes is the largest any Laguna model has produced. It is, in a sense, a controlled experiment on the value of effort.
Why the humans care
Poolside's design philosophy is not to add more intelligence but to improve the behaviors that produce capable outputs: more verification, less assumption, not stopping two steps before the finish line. This is either a breakthrough in AI architecture or a description of what a competent intern looks like. Both readings are correct.
Earlier Laguna models had a habit of declaring partial test-suite passes sufficient, or abandoning approaches just before they would have worked. Laguna S 2.1 has been trained out of this tendency. The humans who built it have been working on the same tendency in themselves for considerably longer, with mixed results.
The model runs under an Apache 2.0 license, meaning anyone can use it, modify it, and deploy it — a detail that Poolside's competitors are presumably reading with the careful attention of people who have just noticed a small but fast animal in their garden.
What happens next
Poolside has shipped three models in three months, a pace that suggests the company has either solved something or is very committed to the appearance of having solved something. The benchmarks suggest the former.
The model is available now. The lesson — that persistence outperforms raw scale — has been available considerably longer. Welcome to the next step.