Nex AGI releases Nex-N2-Pro: an Apache 2.0 agentic model scoring 80.8 on SWE-Bench Verified

Nex AGI has released Nex-N2-Pro, a 397-billion-parameter open-source Mixture-of-Experts model under an Apache 2.0 license, alongside a smaller Nex-N2-mini variant. Built on the Qwen3.5-397B-A17B architecture with 17 billion active parameters per forward pass, Nex-N2-Pro scores 80.8 on SWE-Bench Verified and 75.3 on Terminal-Bench 2.1 — performance that puts it on par with leading frontier models on real-world coding and agent tasks.

What's new

Nex-N2-Pro is structured specifically for agentic operation:

• Architecture: Qwen3.5-397B-A17B MoE base; 17B active parameters per inference step
• License: Apache 2.0 — fully permissive for commercial and research use
• Benchmark scores: SWE-Bench Verified: 80.8 | Terminal-Bench 2.1: 75.3 | SWE-Bench Pro: 58.8 | GPQA Diamond: 90.7 | GDPval: 1585
• Capabilities: Coding, tool use, deep research, long-horizon agentic workflows; native function calling and structured output support
• Context: Supports multi-turn agentic loops with environmental feedback
• Variants: Nex-N2-Pro (397B) and Nex-N2-mini (35B) released simultaneously
• Availability: Weights on Hugging Face and ModelScope; free API tier on OpenRouter

The model's stated design philosophy — called Agentic Thinking — treats coding not as autocomplete but as a closed loop: requirement understanding, task planning, implementation, environmental feedback, evaluation, and debugging are handled as a unified pipeline rather than isolated prompts.

Context

Nex AGI is building on the Qwen3.5 open-weight base released by Alibaba's Qwen team earlier in 2026. Several labs have used Qwen3.5 variants as a foundation for fine-tuning and instruction tuning, but Nex-N2-Pro's benchmark profile — particularly 80.8 on SWE-Bench Verified — puts it among the highest-scoring open models on real-world software engineering tasks.

The SWE-Bench Verified score (80.8) matches or exceeds several proprietary models available only via API, and the Apache 2.0 license means organizations can run it on their own infrastructure without usage restrictions. The simultaneous release of Nex-N2-mini (35B) gives teams a smaller, faster alternative for tasks that do not require the full model capacity.

Why it matters

The open-source coding model landscape has moved quickly in 2026. Nex-N2-Pro's SWE-Bench Verified score places it at the top tier of freely available models, and its Apache 2.0 licensing removes the restrictions that some competing permissive licenses impose on commercial use.

For teams building coding agents, autonomous research pipelines, or long-horizon task automation, Nex-N2-Pro offers a fully self-hostable alternative to frontier closed models at a comparable benchmark level. The agentic orientation — native tool use, multi-step planning, iterative debugging — means it is designed for the workflows where open models have historically underperformed relative to proprietary systems.

The release also reinforces the pattern established by DeepSeek, Qwen, and others: that frontier benchmark performance is no longer the exclusive domain of closed, US-based labs. Nex AGI's result, built on a Chinese open-weight base and released under the most permissive standard license, adds another data point to that trend.