Open-Weight Nex-N2.5-Pro achieves 82.2 score on OSWorld

Nex-AGI releases their open multimodal model Nex-N2.5-Pro (397B params) under Apache 2.0 license. Their model scored 82.2 on OSWorld-Verified (an AI benchmark involving interacting with real GUIs/browsers), 82.7 on Terminal-Bench 2.1, 61.2 on SWE-Bench Pro, 68.5 on Toolathlon Verified. Model weights are publicly available on Hugging Face. Available through OpenRouter.

This is very exciting news as an 82.2 score is amazing for an open-weight model. Now we can finally solve some GUI control tasks! But remember, this model has 397 billion parameters, so you probably can’t run this model yourself without a cluster costing tens of thousands of dollars. With OpenRouter, you can avoid this problem, but keep in mind that you still need to consider tokens when using APIs.