Nex AGI Releases Nex-N2-Pro: 17B Active Parameter MoE Model with 262K Context Window

Nex AGI has released Nex-N2-Pro, a mixture-of-experts model with 17 billion active parameters from a total of 397 billion parameters. The model is currently available for free through OpenRouter.

Technical Specifications

Nex-N2-Pro is built on the Qwen3.5 architecture and offers a 262,144 token context window. The model accepts text and image input and produces text output, classifying it as multimodal. According to Nex AGI, the model supports reasoning capabilities, function calling, and structured outputs.

The mixture-of-experts architecture activates 17B parameters per forward pass while maintaining a total parameter count of 397B, a design choice aimed at balancing computational efficiency with model capacity.

Target Use Cases

Nex AGI positions Nex-N2-Pro for coding tasks, tool use, deep research, and long-horizon agentic workflows. The company claims the model unifies planning, code implementation, debugging, and iteration into a single execution loop, though independent benchmarks have not yet verified these capabilities.

The model's release date is listed as June 8, 2026 on OpenRouter, though this appears to be an error as that date is in the future. The actual release date has not been confirmed.

Availability and Pricing

Nex-N2-Pro is offered as a free model through OpenRouter, with no disclosed pricing for input or output tokens. OpenRouter routes requests to available providers and includes fallback options to maximize uptime.

No benchmark scores, training data cutoff dates, or quantitative performance metrics have been published by Nex AGI at this time.

What This Means

The free availability of a 397B parameter MoE model with a 262K context window represents significant compute resources being offered at no cost, though the sustainability of this pricing model remains unclear. The mixture-of-experts architecture's 17B active parameters suggests inference costs comparable to smaller dense models while potentially accessing the knowledge of a much larger model. However, without published benchmarks or independent evaluations, actual performance relative to established models like GPT-4, Claude, or Qwen's own models cannot be assessed.