Nex Agi: Nex-N2-Pro
Open-weight agentic coding model with 397B MoE (17B active), ~256K context, and free pricing. Scores 80.8 on SWE-Bench Verified and 90.7 on GPQA Diamond, though benchmarks are self-reported.
Specifications
| Attribute | Value |
|---|---|
| Lab | Nex Agi |
| Tags | Agentic Open Weight Image Input |
| Release Date | 2026-06 |
| Context Window | 262,144 tokens |
| Input Price / 1M | $0.00 |
| Output Price / 1M | $0.00 |
| Input Modalities | Text, Image |
| Output Modalities | Text |
Strengths
- 80.8 SWE-Bench Verified (self-reported)
- 90.7 GPQA Diamond (self-reported)
- Free input/output pricing per million tokens
- ~256K context window for long-horizon tasks
- Open-weight Apache 2.0 license for self-hosting
Weaknesses
- 33.6 DeepSWE trails GPT-5.5 (70) and Opus 4.7 (54)
- Self-reported benchmarks; no independent verification
- Limited API providers: OpenRouter and SiliconFlow only
- Requires 16× H100 GPUs for full-precision inference
Best For
- Agentic coding and tool-use workflows
- Cost-efficient multimodal reasoning
- Self-hosted AI agent deployments
- Fine-tuning and customization projects
In Depth: Nex-N2-Pro
Benchmark Performance
Nex AGI: Nex-N2-Pro agentic coding model earns a 7.8/10 LMRank and #86 overall. On SWE-Bench Verified it scores 80.8, within striking distance of Claude Opus 4.7 (rank #3, 9.6/10) despite a fraction of the cost.
Self-reported benchmarks show Nex-N2-Pro at 83.7 on BrowseComp, 90.7 on GPQA Diamond, and 75.3 on Terminal-Bench 2.1 (highest among open models per Nex AGI). However, DeepSWE performance (33.6) lags GPT-5.5 (70) and Opus 4.7 (54), and all scores await independent verification.
Pricing & Value
Nex AGI: Nex-N2-Pro pricing is $0.00 per million tokens for both input and output completely free on the initial OpenRouter tier.
At $0.00/M tokens, Nex-N2-Pro costs $0.00 per LMRank point, versus $15.00/M input for Claude Opus 4.8 (rank #2, 9.7/10) at roughly $1.55 per point, or GPT-5.5 Pro at $30.00/M input ($3.16 per point). Even after the free period, paid rates on OpenRouter ($0.25 in / $1.00 out) remain far below closed peers.
Who Should Use This
Nex AGI: Nex-N2-Pro reasoning model suits developers who need free, open-weight access for agentic tasks. Researchers fine-tuning on coding workflows will benefit from the Apache 2.0 license and ~256K context.
- Agentic coding teams: free pricing with strong SWE-Bench, but DeepSWE lags closed models
- Self-hosters: open-weight Apache 2.0 license, but requires 16× H100 GPUs
- Budget-conscious researchers: zero-cost API access, but limited provider support
- Not for teams needing DeepSWE scores above 70 or verified benchmarks
Release & Version History
Nex AGI: Nex-N2-Pro launched in June 2026 as the successor to Nex-N1 (December 2025). The model builds on the Qwen3.5-397B-A17B base with a Mixture-of-Experts architecture.
Nex AGI released Nex-N1 in December 2025, establishing its open-weight MoE approach. Six months later, Nex-N2-Pro arrived with the Agentic Thinking framework (Adaptive + Coherent Thinking). A smaller sibling, Nex-N2-Mini, launched alongside. The model has 397B total parameters (17B active) and supports image input, tool calling, and structured outputs.