Nex-N2.5-Pro

Nex N2.5 Pro

Quick answer: Nex-N2.5-Pro is Nex-AGI's mid-tier open-weight agentic model in the Nex-N2.5 family (mini/Pro/Max), built for long-horizon computer-use, web-browsing, and coding agent tasks. It scores 82.7% on Terminal-Bench 2.1, 61.2% on SWE-Bench Pro, and 82.2% on OSWorld-Verified, and is available both as open weights on Hugging Face and hosted via OpenRouter.

At a Glance

Where Nex N2.5 Pro leads

• 82.2% OSWorld-Verified and 56.4% OSWorld-2 — strong computer-use grounding for an open-weight model
• 87.4% OSWorld-G — leads several closed frontier rivals shown on its own card (e.g. Claude Opus 5 at 76.8%)
• 68.2% Vision2Web — competitive web-GUI-agent grounding
• Open weights + hosted OpenRouter access — flexible deployment

Where it lags

• 61.2% SWE-Bench Pro — behind Claude Opus 5 (79.2%) and DeepSeek-V4-Pro-0813 (67.7%)
• 44.2% AutomationBench v1.0.6 — trails most frontier rivals on its own comparison table
• 38.2% SWE-MM — weakest shown among compared models on multimodal SWE tasks

Best for: teams wanting an open-weight, self-hostable agent model with strong computer-use/web-browsing grounding rather than best-in-class raw coding accuracy.

What Nex N2.5 Pro Is

Nex-N2.5-Pro is the mid-size model in Nex-AGI's second-generation N2.5 agentic family, sitting between Nex-N2.5-mini and the 1.6T-parameter Nex-N2.5-Max. It builds on the multimodal foundations of the earlier Nex-N2 line, with targeted improvements in computer use, web browsing, and visually-grounded agentic capability — treating vision not just as an input modality but as the feedback channel an agent uses to verify its own actions in a browser or desktop environment.

Coding evaluations use Nex-AGI's own NexAU harness; computer-use and browser-use benchmarks (OSWorld, WebTest, WebArena) run through a NexCUA harness with grounding coordinates normalized to a 0–1000 scale (soon to be open-sourced). The model supports a reasoning_effort parameter (none/medium/high) to trade inference cost for deliberation, and ships with SGLang deployment recipes for single-node 8×H100 serving.

Specifications

Pricing

Open-weight model; self-hosting cost only, or use via OpenRouter at provider-set rates.

Public Benchmark Scores

Scores above are reported by Nex-AGI and shown for context. They are not Benchgen measurements. Coding tasks use the NexAU harness; computer-use/browser tasks use the NexCUA harness. WebTest, WebArena-Verified, OSWorld-G, Vision2Web, and SWE-MM scores are documented in the accompanying bucket-B benchmark configs added alongside this model rather than repeated here to avoid duplicate unsourced rows before those pages exist; see each benchmark's own leaderboard entry for this model's score.

Nex N2.5 Pro vs Alternatives

*DeepSeek-V4-Pro-0813 score as reported on the Nex-N2.5-Pro card; see DeepSeek's own card for a possibly different harness/date.

Nex-N2.5-Pro trades top-tier coding accuracy for stronger computer-use and browsing grounding at open-weight cost — a fit for teams building browser/desktop agents who want to self-host rather than pay closed-API prices, at the cost of some ground on pure SWE-bench-style coding tasks versus Claude Opus 5 or DeepSeek V4 Pro.

How Nex N2.5 Pro Performs on Real Agent Tasks

Nex-N2.5-Pro's benchmark spread signals a model tuned specifically for the "verify with vision" loop of computer-use agents: it leads or nearly matches frontier closed models on OSWorld-G (87.4%, ahead of Claude Opus 5's 76.8%) and stays competitive on OSWorld-Verified, while trailing further on text-only coding benchmarks like SWE-Bench Pro and AutomationBench. That pattern is consistent with a lab explicitly optimizing for browser/desktop agent reliability over raw code-generation accuracy — useful context for anyone picking a model based on a single averaged score rather than the task-specific breakdown.

Use Nex N2.5 Pro via API

from openai import OpenAI  # OpenRouter exposes an OpenAI-compatible API

client = OpenAI(api_key="YOUR_OPENROUTER_KEY", base_url="https://openrouter.ai/api/v1")

response = client.chat.completions.create(
    model="nex-agi/nex-n2.5-pro",
    messages=[{"role": "user", "content": "Book a flight using this browser session..."}],
)
print(response.choices[0].message.content)

Related Models & Benchmarks

• Compare: DeepSeek V4 Pro · Nex N2.5 Max
• Benchmarks: Terminal-Bench 2.1 · OSWorld-Verified · SWE-Bench Pro

Frequently Asked Questions

Specs and scores sourced from Nex-AGI's official Hugging Face model card. Last updated 2026-09-15.