Nex-N2-Pro Unifies Reasoning, Tools, And Code For Agentic Workflows
Nex-N2-Pro is a new open-source AI model designed to handle complex, real-world agentic tasks. It unifies reasoning, tool use, and code execution into a single continuous workflow called Agentic Thinking. The model is released in two sizes to balance performance and resource needs.
The team at Nex-agi built Nex-N2-Pro and its smaller sibling Nex-N2-mini to push AI agents beyond simple dialogue. Both variants are post-trained on the Qwen3.5 architecture and optimized for long-horizon productivity scenarios. They aim to let the model decide when to think deeply and when to act quickly.
Features for agentic workflows
Key Features
- Adaptive Thinking for efficient deep reasoning.
- Coherent Thinking across all agentic tasks.
- Strong coding scores on Terminal-Bench 2.1.
- Excels at long-horizon tasks like GDPval.
- Built on Qwen3.5 architecture with two sizes.
- Supports function calling and reasoning traces.
- Requires custom sglang fork for best performance.
This tool suits developers and small teams needing a private AI agent for coding, research, or automation. Running locally on high-end GPUs keeps data in-house and avoids cloud dependencies. The Pro variant needs serious hardware, while the mini version offers a lighter testbed.
Deployment and known limitations
Deploying the Pro model requires two 8× H100 nodes, making it impractical for home labs. The mini variant runs on a single 2× H100 server. Docker images and recommended sampling parameters (temperature 0.7, top_p 0.95) are provided for consistent quality.
“Rather than treating reasoning, tool use, and environment execution as separate capabilities, Nex-N2 unifies them through an Agentic Thinking framework.” — Source: Hugging Face
Project Links
