Nex-N2.5-Pro: Where a 397B Agent Model Earns Its Keep

Nex-N2.5-Pro beats its smaller sibling on every benchmark Nex AGI published. That is expected and, on its own, useless: a bigger model in the same family usually wins. The useful question is where the gap is wide and where it is barely there, because that tells you which tasks are worth paying for Pro and which are not.

Nex's own table answers that question more clearly than most launch tables do, and the pattern is consistent enough to build a routing rule on.

What Pro is

Pro is the middle tier of the N2.5 family, released on 8 September 2026 alongside Mini and Max. Its Hugging Face repository lists 397B parameters under a qwen3_5_moe architecture tag, with BF16 and FP8 weights, and an Apache 2.0 licence. Like Mini, it takes images as well as text and is aimed at computer use, browsing and coding agents. The card does not state an active parameter count.

Nex's reference deployment is a single node with eight H100s, using its patched SGLang build with the qwen3 reasoning parser and the qwen3_coder tool-call parser. The same reasoning_effort control as Mini applies: none, medium by default, or high.

Apache 2.0 at this size is worth noting. You can fine-tune it, serve it to customers and redistribute it without a revenue threshold or a separate agreement, which is not true of every open model in this class. Open weights vs open source explains why the distinction matters most once a product succeeds.

Where Pro pulls away from Mini

All figures here are Nex's own, from the model cards. The largest gaps are on tasks that run for many steps:

• OSWorld-2 — 56.4 against Mini's 30.5
• DeepSWE v1.1 — 55.8 against 36.1
• SWE-Bench Pro — 61.2 against 43.8
• Toolathlon Verified — 68.5 against 54.6
• OSWorld-Verified — 82.2 against 71.2
• Terminal-Bench 2.1 — 82.7 against 73.4

The smallest gaps are on tasks closer to perception or a single page of interaction:

• OSWorld-G, a GUI grounding test — 87.4 against 82.9
• WebTest — 52.8 against 48.6
• BrowseComp — 89.7 against 83.4
• OmniDoc, document parsing — 92.2 against 89.7

Read together, Mini can already see the screen and find the button almost as well as Pro. What it cannot do as well is keep a long task on track. That is the property you are paying for.

What that means for routing

A computer-use agent mixes both kinds of work. Locating an element, reading a form, extracting a table from a screenshot — those are the grounding tasks where Mini nearly matches Pro. Planning a twenty-step workflow, recovering after a page did something unexpected, or fixing a failing test across several files — those are where Pro's lead is widest.

That suggests two practical shapes. One is a planner-executor split: Pro plans and recovers, Mini executes individual grounded steps. The other is escalation: run Mini by default and hand the task to Pro when it fails or stalls. Model routing and fallbacks covers both, and the right choice depends on whether your failures are expensive or merely slow.

For pure coding agents the case is simpler. The gaps on SWE-Bench Pro and DeepSWE are large enough that Mini is unlikely to be a substitute for multi-file repository work. SWE-bench explained covers what that benchmark family actually measures.

Cost: three times Mini, still cheap

At the time of writing OpenRouter lists Pro at $0.075 per million input tokens and $0.25 per million output, with cached input at $0.015. That is three times Mini's input price and two and a half times its output price.

In absolute terms both are inexpensive, and that changes the usual calculation. When the cheaper model is a fraction of a cent per task, the deciding cost is not tokens but failed runs: the retries, the human who has to finish the job, the wall-clock time. If Pro completes long tasks Mini abandons, it is cheaper per completed task even at three times the token price.

Self-hosting is a different matter. Eight H100s against two is a fourfold hardware difference before you count utilisation, and that is where the Mini-first routing pattern pays for itself.

The harness caveat

Nex ran its coding evaluations in its own NexAU harness and computer-use evaluations in NexCUA, at temperature 0.7, top-p 0.95 and top-k 40. Comparison columns mix leaderboard figures, provider reports and Nex's own runs. That is normal for a launch table, and it means absolute numbers should be discounted — the reported 1,628 on GDPval-AA v2, for instance, is a vendor figure on an evaluation where harness and grading details matter a great deal.

The within-family comparison above is more robust than the cross-vendor one, because Mini and Pro were run the same way. If Nex's harness flatters both, it flatters both roughly equally.

How to test it on your own agents

1. Pick twenty tasks your agent actually runs, split between short grounded steps and long multi-step jobs.
2. Run each on Pro and on Nex-N2.5-Mini with the same harness, prompts and recommended sampling settings.
3. Record completion, turns taken and total tokens, separately for the short and long groups.
4. Check tool calling on your provider before anything else; a parser misconfiguration looks exactly like a weak model.

If Pro's advantage shows up only in the long group, as the vendor table suggests it should, you have your routing rule.