Nex-N2 Pro and Mini: Nex AGI's Models Explained

A 397-billion-parameter open-source model claiming 80.8 on SWE-Bench Verified dropped in June 2026, and most people never heard about it. That model is Nex-N2 Pro from Nex AGI, and its smaller sibling Nex-N2 Mini followed a couple of weeks later. No launch event, no viral thread. Just two Apache 2.0 releases on Hugging Face and a model card with numbers bold enough to raise eyebrows. This guide covers what the Nex N2 Pro and Mini actually are, what they cost, and which one is worth your time.

What Are the Nex-N2 Pro and Mini Models?

Nex AGI is a newer lab that has taken a specific bet: instead of building a general chat model and bolting agent skills on top, the Nex-N2 series treats agentic work as the main event. The lab calls its approach an "Agentic Thinking" framework — one loop that connects understanding the task, planning it, writing code, reading feedback from the environment, evaluating results, and iterating until the job is done. Most models do some of that. Nex-N2 was trained to do all of it as a single continuous behavior.

The practical result is a pair of models that feel built for tool use and multi-step coding rather than casual conversation. Both accept images as well as text, which matters more than it sounds — an agent that can read a screenshot of a stack trace or a UI mockup skips a lot of transcription friction.

The family line is easy to remember. Pro is the workhorse: 397B total parameters, about 17B active per token thanks to its mixture-of-experts design. Mini is the same idea at pocket size: around 35B total, 3B active. Same 262,144-token context window on both, which comfortably fits a mid-sized codebase or a few hundred pages of documents.

How Do Nex-N2 Pro and Mini Compare on Specs and Price?

Here are the two models side by side, with the per-token list pricing you'll see on model marketplaces. (On CoreAI you don't deal with per-token math at all — one subscription covers both — but the list prices tell you where each model sits in the market.)

Read that pricing row twice. Pro at $0.25 in / $1.00 out per million tokens is already cheap for a frontier-adjacent agentic model. Mini at $0.025 in / $0.10 out is practically a rounding error — it sits in the same cost neighborhood as the models we covered in our budget AI models roundup, and undercuts most of them.

Are the Nex-N2 Benchmark Claims Real?

Here's where honesty is required. On Nex AGI's own reported numbers, Nex-N2-Pro scores 80.8 on SWE-Bench Verified, 90.7 on GPQA Diamond, 94.0 on IFEval, and 58.8 on SWE-Bench Pro. Those figures would put an open-source model in the same conversation as closed frontier systems, and they're published on the official Hugging Face model card.

The caveat: as of mid-July 2026, none of these numbers have been independently reproduced at scale, and early hands-on reception has been more measured than the model card. That's not an accusation — self-reported benchmarks from young labs often land a few points high once the community re-runs them. It's just a reason to test the model on your own work before reorganizing your stack around it.

The good news is that testing it takes about a minute. Open CoreAI's Compare view, put Nex-N2-Pro next to whatever model you currently trust, paste in a real task from your week, and watch both answers stream side by side. Benchmarks argue; side-by-side outputs settle.

One more genuinely useful detail: Apache 2.0 licensing means teams can self-host, fine-tune, and ship commercial products on either model with no strings attached. For companies that got burned by restrictive "open" licenses, that's a real differentiator, and the weights are all public on the Nex-N2 GitHub.

When Should You Pick Pro, and When Is Mini Enough?

Our take, after running both against day-to-day tasks:

Pick Nex-N2-Pro when the task has more than three steps. Multi-file code changes, debugging sessions where the model needs to reason about test output, research tasks that involve reading a long document and producing structured work. Pro's larger expert pool shows up as steadier planning — it loses the thread less often on step six. With reasoning enabled (see our guide to when thinking mode pays off), it's a legitimately strong coding companion.

Pick Nex-N2-Mini for volume work. Summarization, extraction, classification, first-draft code snippets, reformatting — anything you'd run hundreds of times. At a tenth of Pro's price and with the same 262K context window, Mini is the model you point a firehose at. It will occasionally fumble a subtle instruction that Pro catches, which is the trade you're making.

Skip both when you need heavy multimodal output or maximum polish. Neither model generates images, and for prose where tone really matters, the big-name flagships still write with more finesse. Nex-N2 is a tool-user, not a stylist.

The honest summary: Mini is the easier recommendation. Budget-tier pricing with a 262K window and image input is a rare combination, and the downside risk at that price is basically zero. Pro is promising but needs its benchmark story confirmed — try it on your own prompts before you commit.

Frequently Asked Questions

Is Nex-N2 Pro open source?

Yes. Both Nex-N2-Pro and Nex-N2-Mini are released under the Apache 2.0 license, with weights available on Hugging Face. That permits commercial use, modification, and self-hosting without meaningful restrictions.

What context window do the Nex-N2 models have?

Both models share a 262,144-token context window — enough for roughly 500 pages of text or a mid-sized code repository in a single conversation. Pro and Mini are identical on this spec, so choosing Mini doesn't cost you any context.

Can Nex-N2 models see images?

Yes. Both Pro and Mini accept text and image input and produce text output. That covers screenshots, diagrams, and photographed documents, but neither model generates images itself.

How much does Nex-N2 Mini cost compared to Pro?

Mini lists at $0.025 per million input tokens and $0.10 per million output tokens — almost exactly one tenth of Pro's $0.25/$1.00. On CoreAI, both are included in one subscription, so you can switch between them freely mid-conversation.

Where can I try Nex-N2 Pro without setting anything up?

Both models are live in the CoreAI model library on iOS, Android, and the web. No downloads, no configuration — pick the model, type, and compare it against 300+ others if you want a second opinion.