Nex-N2.5

Nex-AGI's second-generation agentic family, released September 8, 2026 in three sizes. mini and Pro continue the multimodal Nex-N2 line with focused gains in computer use, web browsing, and visually grounded agent work: their configs are Qwen3.5 MoE shapes (mini 256 experts, 8 active; Pro 60 layers, 512 experts, 10 active; 262K context), so they are post-trained on Qwen3.5 bases. Max is new: a text-only model on the DeepSeek-V4 architecture (61 layers, 384 routed experts, 6 active, 1M context) that Nex-AGI describes as its first "systematic post-training at trillion-parameter scale," which places it on DeepSeek-V4-Pro. All three ship under Apache 2.0 on HuggingFace and ModelScope, with hosted access on OpenRouter.

Self-reported (NexAU harness for coding, NexCUA for computer use): Max reaches 86.1 on Terminal-Bench 2.1, 65.7 on SWE-Bench Pro, 65.6 on DeepSWE v1.1, 50.2 on AutomationBench (level with Claude Opus 5's 50.3), 74.7 on Toolathlon Verified, 1713 GDPval-AA Elo, and 92.6 on BrowseComp, the top figure in its comparison table; Pro scores 82.2 on OSWorld-Verified, 87.4 on OSWorld-G, and 89.7 on BrowseComp. The NexCUA computer-use harness is promised as open source. Parameter counts are not stated on the cards; the weights inherit the bases' sizes. Not yet scored by Artificial Analysis at filing.