Nex AGI: Nex-N2.5-Pro

Nex-N2.5-Pro is the mid-tier member of Nex-AGI's Nex-N2.5 family of agentic models, positioned between Nex-N2.5-mini and the trillion-parameter text-only Nex-N2.5-Max. It extends the multimodal Nex-N2 line with stronger computer use, web browsing, and visually grounded agent workflows, treating vision as an active control and verification interface rather than passive input. The model is built for long-horizon tasks in real environments, including terminal coding, browser automation, and knowledge work with continuous action and self-correction from visual feedback.

Architecturally, Nex-N2.5-Pro is a 397-billion-parameter mixture-of-experts model with a Qwen3-style multimodal stack: 60 transformer layers, 4,096 hidden size, 512 routed experts with 10 active per token, hybrid linear and full attention, and a 27-layer vision encoder for image understanding. It supports a native 262,144-token context, optional reasoning modes via reasoning_effort, and OpenAI-compatible tool calling when served with Nex-AGI's recommended SGLang configuration. Weights are released under Apache 2.0 on Hugging Face and ModelScope, with hosted access on OpenRouter.

On Nex-AGI's published evaluation suite, Nex-N2.5-Pro reaches 82.7% on Terminal-Bench 2.1, 61.2% on SWE-Bench Pro, 55.8% on DeepSWE v1.1, 68.5% on Toolathlon Verified, 89.7% on BrowseComp, and a GDPval-AA v2 rating of 1628, with 82.2% on OSWorld-Verified in multimodal computer-use settings. Reference self-hosting uses tensor parallelism across eight H100 GPUs on a single node.