Nex-AGI: Six Agentic Models, One of Which Fits Your GPU

Six models from Nex-AGI just landed in the model library. They are worth a closer look than the usual launch-week entry, because the lab behind them is doing something structurally different from most of the names in that catalogue — and because exactly one of the six will ever load on a card you can buy.

Here is what they are, where the numbers come from, and what it actually takes to hold each one.

Who Nex-AGI is

Not a startup in the usual sense. Nex-AGI describes itself as "an innovation alliance initiated by the Shanghai Innovation Institute" — a consortium rather than a company, publishing from nex.sii.edu.cn and nex-agi.com.

The shape of its GitHub organisation tells you what kind of outfit it is. Alongside the weights sit a forked sglang, a forked llama.cpp, a post-training framework (NexRL, described as an "ultra-loosely-coupled LLM post-training framework"), an agent framework (NexAU), a deep-research agent (NexDR), and a Python debugging toolkit. That is a full stack wrapped around inference, post-training and agents.

Conspicuously missing from it: anything for pretraining.

They don't pretrain. They post-train.

This is the fact that explains every other fact about these models. Every Nex release takes someone else's open base weights and post-trains them for agentic work — tool calling, terminal use, long-horizon autonomy — then publishes the result under its own name.

That has a direct consequence for anyone sizing hardware, and it is the reason these records were straightforward to add correctly: a post-train is the base model's weights at the base model's size. Unlike a distillation — which produces a genuinely smaller model — post-training changes behaviour, not shape. Nex-N2-Pro needs exactly what Qwen3.5-397B-A17B needs. Not approximately. Exactly.

It also means the interesting claim these models make is not "we built a better brain." It is "we taught an existing brain to use a computer." Nex calls the method Agentic Thinking, and splits it in two:

• Adaptive Thinking — the model decides for itself when to reason and how deeply, "executing simple actions quickly while reasoning thoroughly on critical decisions." In practice this is exposed as a reasoning_effort parameter with three settings: none, medium (the default, which hands the decision to the model) and high.
• Coherent Thinking — one consistent reasoning style carried across chat, tool calls, terminal work and vision, so behaviour does not fracture when the task changes shape.

Three generations in about a year

Nex-N1 set the pattern: take open bases across a range of sizes — from internlm3-8B up to DeepSeek V3.1 at 671B — and post-train each for agent work. The flagship, DeepSeek-V3.1-Nex-N1, scores 80.2 on τ²-Bench, 70.6 on SWE-bench Verified and 65.3 on BFCL v4.

Nex-N2 moved to the Qwen3.5 series and got substantially better. Nex-N2-Pro reaches 75.3 on Terminal-Bench 2.1, 80.8 on SWE-bench Verified, 90.7 on GPQA Diamond and 1585 on GDPval. It ships under Apache 2.0 — a genuinely permissive licence on a 397B model.

Nex-N2.5 added a third, much larger size and shifted the emphasis to vision. Not vision as an input mode — vision as the thing an agent looks at to check whether its last action worked.

Two things stand out. The 35B mini gained more than twelve points on Terminal-Bench between N2 and N2.5 without changing size at all — post-training alone. And Max's 92.6 on BrowseComp is the top score in Nex's own comparison table, ahead of the frontier proprietary models it lists.

The SWE-Bench Pro column also drops between N2-mini (50.2) and N2.5-mini (43.8). Nex does not explain that, and the two generations were scored in different tables; treat cross-generation comparisons on that row with suspicion.

What they cost to hold

This is where the site's actual job begins. Every figure below is computed by the same VRAM function that drives the rest of the catalogue — weights at Q4_K_M, plus the KV cache at an 8K context, plus 0.8 GB of runtime overhead.

Nex's own serving instructions corroborate every row: 2× H100 for the mini, 8× H100 for Pro, and 16× H200 across two nodes for Max. Five of six models are datacenter hardware. That is not a criticism — it is the honest shape of frontier open weights in 2026 — but it is the thing a launch post usually leaves out.

The one that fits

The 35B mini is the whole local story here, and it is a good one.

At Q4_K_M it needs 23.8 GB with an 8K context. On a 24 GB RTX 4090 or 3090 that is genuinely tight — it loads, but you are not stretching the context far. On a 32 GB RTX 5090 it is comfortable. Our VRAM calculator will show you the curve for your own context length, and the GPU checker will give you a verdict for your specific card.

What makes it interesting is the sparsity. The model holds 35B parameters but routes only 3B per token, so decode reads about 1.81 GB per token instead of 21.1 GB. That is the difference between a 35B model that crawls and one that streams: our modelled estimate puts it at roughly 111–212 tok/s on a 4090 and 139–289 tok/s on a 5090 — ranges, not point values, because these are modelled figures rather than measurements. It runs at small-model speed while carrying a mid-size model's knowledge.

This is the single most misunderstood thing about Mixture-of-Experts models, and it is worth stating plainly: you pay for all 35B in memory and only 3B in speed. The memory bill is not negotiable — every expert has to be resident, because the router may call any of them on the next token. The speed bill is what sparsity buys you.

The 671B model has a cheaper context than the 35B one

Look again at the KV column above. The 671B Nex-N1 carries a 0.58 GB cache at 8K tokens. The 35B mini carries 1.88 GB — more than three times as much, on a model twenty times smaller.

That is not an error. It is Multi-head Latent Attention.

The DeepSeek lineage — V3, V3.1, V4, and therefore both Nex-N1 and Nex-N2.5-Max — compresses each token's key/value pair into a low-rank latent vector rather than storing it per attention head. The cache stops scaling with head count and starts scaling with a much smaller compressed rank. Nex's own launch command for Max confirms the lineage: it selects --attention-backend dsv4 and a deepseek-r1 reasoning parser, while the Qwen3.5-based Pro and mini use qwen3.

The practical upshot: on the DeepSeek-derived models, long context is nearly free relative to the weights. On the Qwen-derived ones it is not. If you were choosing between them for a long-context workload, that ratio matters more than the parameter count.

Vision as an action interface

The N2.5 pitch is worth quoting directly, because it is a sharper claim than "now with images":

> Vision is therefore no longer merely an input modality; it has become a critical interface through which an agent perceives its environment, verifies outcomes, and moves a task forward.

The benchmarks Nex leads with follow from that. OSWorld-Verified and OSWorld-2 measure whether a model can operate a desktop; BrowseComp measures whether it can navigate the web to find something. N2.5-Pro's 82.2 on OSWorld-Verified sits within about a point of the proprietary frontier models in the same table.

Whether that generalises past the benchmark is exactly the sort of claim that needs independent replication, and it does not have it yet. But the framing — screenshots as feedback rather than input — is the direction the whole agentic field has been moving, and it is the reason these models are more than another set of Qwen fine-tunes.

What we are less sure about

Three things, stated plainly, because every number on this site carries a confidence label and these are the ones carrying the weaker ones.

The N2.5 parameter counts are inferred, not published. Nex states the N2 figures outright. For N2.5-mini and N2.5-Pro it says only that they "continue to build on the multimodal foundations of Nex-N2." We have recorded them at the N2 sizes and marked those variants preview rather than verified. The serving configurations back the inference — 2× H100 and 8× H100 are the right shapes for 35B and 397B respectively, and wrong for anything else in the family — but an inference is not a spec sheet.

The N2.5 licence is not published. The card says the weights "will be released as open source" without naming terms. Nex-N2 is confirmed Apache 2.0 and is recorded that way. The N2.5 records say open-weight, terms unpublished, which is the conservative direction: it tells you to go and read the licence rather than assuring you that you may ship a product on it.

Nothing here has been independently benchmarked. Every score in this post comes from Nex's own tables. Nex is transparent about its methodology — sampling parameters, evaluation harness, which competitor scores were lifted from official leaderboards versus run in-house — which is more than many labs disclose. It is still a vendor reporting on itself.

Why they are in the library

Because the questions people bring to this site are factual ones, and they now have answers.

Can I run Nex-N2-mini on a 4090? Yes, tightly, at roughly 111–212 tok/s. What does Nex-N2.5-Max need? 967 GB, which is sixteen H200s. Is Nex-N2-Pro safe to build a product on? Yes — Apache 2.0. Is N2.5-Pro? Unknown, and that is the honest answer until Nex publishes terms.

Those answers do not depend on whether the models are any good, or on whether the alliance behind them is still shipping in a year. They are arithmetic and licence text, and they are the same two things that decide every model on this site.

You can compare all six against anything else in the catalogue with the model comparison tool, check them against your own hardware in the analyzer, or read how the VRAM and throughput numbers are derived.