The Sequence AI of the Week #903: Laguna, the 118 Billion Parameters that Walks Into a Trillion-Parameter Bar

Plot every open-weight model's parameter count against its Terminal-Bench 2.1 score and you get the graph everyone expects: a smooth upward slope, bigger models doing better, no surprises. Then Laguna S 2.1 shows up at 118 billion parameters with a score of 70.2%, sitting well above where that trend line says it should be.

For context, DeepSeek-V4-Pro-Max weighs in at 1.6 trillion parameters — more than 13 times the size — and only manages 64.0. Inkling, at 975 billion parameters, scores 63.8. Nemotron 3 Ultra, at 550 billion, lands at 56.4. So the smallest model in this comparison is beating models that are 5x, 8x, and 13x its size. On DeepSWE, a benchmark that's tougher and less picked-over, the gap turns into a chasm: Laguna hits 40.4 while DeepSeek-V4-Pro-Max manages just 9.0.

A result like this — huge score advantage, tiny fraction of the parameters — is usually the point where you start hunting for a broken benchmark or a contaminated training set. Poolside, the company behind Laguna, seems to have expected that skepticism. They published every trajectory from every trial in the final evaluation run. Anyone can go look at exactly what the model did on each task, not just the aggregate number.

That transparency choice says a lot about how this release was built. It's not a company trying to sneak past scrutiny with a flashy chart. It's a company betting that if you show your work, the numbers will hold up on their own. Poolside is not a household name yet, but results like these — especially paired with open trajectories instead of just marketing slides — tend to change that fast.