Nex-N2 Tested: Open-Source Agentic Model Builds a Full-Stack App Free

Key Takeaways

• Two open-source flavours, currently free. Next N2 Pro is built on a ~400B-parameter Qwen 3.5 MoE base; Next N2 Mini targets lower latency. Both are open-sourced and free to use on hosted providers like OpenRouter — but they're large, so self-hosting needs a multi-GPU cluster, not a single GPU.
• Harness-agnostic. Fahd ran it through the Hermes coding agent on Ubuntu using a generic profile (you can also use a dedicated coder profile). The model isn't tied to any one agent harness.
• The test was deliberately hard. A single prompt asked it to build a Django REST + React banking dashboard and intentionally introduce five real-world bugs spanning dependency management, backend config, database state, auth middleware, and frontend config — then fix them.
• "Loop coding" in action. The run was fully agentic: it planned, generated the repo, wrote its own tests, and looped to fix until the goal verified — 1h 15m, ~144 tool calls, 66K+ tokens on the free hosted model.
• Result: it works. Backend and frontend both launched, demo login worked, and transactions posted correctly — a real full-stack app from one prompt. The honest benchmark read: strong on reasoning, instruction-following and terminal tasks, but frontier proprietary models still lead on the most demanding SWE evals.

Commands & Code Mentioned

# Set the Hermes coding agent (generic profile, no coder profile)
hermes

# After the agent finishes, run the generated app
source venv/bin/activate   # activate the Python virtualenv for the Django backend
npm run dev                # start the React frontend dev server

Exact commands were provided in the agent's own output; the steps above reflect what Fahd ran on camera to launch the generated backend and frontend.