Nex AGI releases Nex-N2-Pro, an open agentic MoE model with 262K context

Nex AGI has released Nex-N2-Pro, a new agentic mixture-of-experts language model aimed at coding, tool use, deep research and long-horizon workflows.

OpenRouter’s listing for Nex AGI: Nex-N2-Pro describes the model as a June 8, 2026 release with a 262K-token context window. The OpenRouter page positions the model for agent-style tasks including coding, tool use, deep research and extended workflows.

Nex AGI’s own product page describes Nex-N2-Pro as a 397B-parameter mixture-of-experts model based on Qwen3.5-397B-A17B. The company says the model is built around what it calls “Agentic Thinking,” a framing for models that can reason through tasks while interacting with tools or longer processes.

The official nex-agi/Nex-N2-Pro model card on Hugging Face says Nex-N2 is being released and open-sourced as an agentic model with Agentic Thinking. The same Hugging Face page lists the model under the Apache-2.0 license, a permissive open-source license that generally allows broad reuse subject to its terms.

What the model is targeting

Across OpenRouter, Nex AGI’s product page and the Hugging Face model card, Nex-N2-Pro is presented as a model for agentic workloads rather than only short-form chat. OpenRouter specifically names coding, tool use, deep research and long-horizon workflows as target use cases.

The model’s reported 262K context window is notable because long-context capacity can matter for tasks such as reviewing large codebases, following multi-step research trails or maintaining state across extended tool-driven sessions. The sources do not, however, independently verify performance in production settings, so the context size should be read as a listed specification rather than proof of reliability on every long-context task.

Nex AGI’s page also reports benchmark results for coding and long-horizon agentic tasks. Those benchmark claims come from the company’s own product materials, and should be interpreted as vendor-reported results unless independently reproduced.

Open availability and distribution

The Hugging Face model card is the clearest source on availability: it identifies nex-agi/Nex-N2-Pro as an official model release and lists the license as Apache-2.0. OpenRouter’s listing separately makes the model accessible through its model routing platform, where developers can compare and call supported models through OpenRouter’s API ecosystem.

Together, the sources indicate that Nex AGI is pursuing both open model distribution and hosted access channels. The Hugging Face release supports users who want to inspect or deploy the model according to the license terms, while OpenRouter provides a third-party route for developers who use its model marketplace.

Why it matters

Nex-N2-Pro enters a growing category of large open or semi-open models designed not just for answering prompts, but for acting across longer task chains. Its combination of a vendor-stated 397B-parameter MoE architecture, a listed 262K-token context window and Apache-2.0 licensing makes it relevant for developers evaluating open agentic models for software engineering, research assistance and tool-using systems.

The practical significance will depend on independent testing, deployment cost, latency, tool-use reliability and how well the model handles real-world long-horizon tasks. For now, the strongest source-backed takeaway is that Nex AGI has published Nex-N2-Pro as an open agentic MoE model, OpenRouter lists it as available with long-context support, and Hugging Face identifies the release as Apache-2.0 licensed.