Nex-AGI Drops Nex-N2.5, a 397B Open Agent That Controls Your Desktop
Nex-AGI releases an open-source agentic model family with vision-driven computer use, three sizes, and benchmark scores rivaling Claude Opus 5.
- Nex-AGI released the Nex-N2.5 family: mini, Pro (397B MoE), and Max (1.6T) under Apache 2.0.
- Focus on long-horizon agents that use vision as a feedback loop for computer and browser control.
- Pro hits 87.4 on OSWorld-G, leading Claude Opus 5 and GPT-5.6 Sol on GUI grounding.
- Weak spots: SWE-MM (38.2) and Job Bench (41.4) trail Claude Opus 5 by wide margins.
- Runs on 8xH100 via a customized SGLang fork with OpenAI-compatible endpoints and function calling.
- Hosted access available on OpenRouter; over 30,000 downloads on Hugging Face already.
Nex-N2.5 brings visual agents to open-weight deployment
Nex-AGI has released the Nex-N2.5 family, a trio of models designed for long-running agent tasks such as controlling browsers, operating desktop software, and executing and debugging code. The middle-tier Nex-N2.5-Pro is a 397-billion-parameter mixture-of-experts model distributed under the Apache 2.0 license.
Nex-N2.5-Pro is available through the OpenRouter API and as downloadable weights from Hugging Face or ModelScope. Apache 2.0 permits commercial use, modification, and redistribution subject to its license and notice requirements.
The release in one screen
- Models: Nex-N2.5-mini, Nex-N2.5-Pro, and Nex-N2.5-Max
- Pro architecture: 397-billion-parameter mixture of experts
- Modalities: Mini and Pro support multimodal tasks; Max uses a text-only foundation
- Target workloads: Browser control, desktop operation, coding, debugging, and tool use
- Pro license: Apache 2.0
- Pro deployment target: One node with eight H100 GPUs
- Serving options: Hosted API or self-hosted SGLang server
A mixture-of-experts model routes each token through a subset of its expert blocks, reducing computation relative to a dense model with the same total parameter count. Self-hosting still requires enough memory to make the full checkpoint available during inference.
Vision closes the action loop
Long-horizon agents must check whether each action produced the intended result before choosing the next step. Nex-N2.5 uses screenshots as an active feedback channel: the model observes the interface, performs an action, inspects the updated screen, and adjusts its plan.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
