Nex-N2.5: Nex-AGI's Mini, Pro, and Trillion-Param Max Agentic Models
Nex-AGI's Nex-N2.5 family (mini, Pro, Max) targets computer use, browsing, and coding agents, with Max built on a 1.6T-param MoE.
What is Nex-N2.5?
Nex-N2.5 is a family of three open-weight models from Nex-AGI, mini, Pro, and Max, built specifically for agentic work: operating computers, browsing the web, and writing and testing code over long, multi-step tasks. Nex-N2.5-Max is the standout of the release, a 1.6-trillion-parameter Mixture-of-Experts (MoE) model, the company’s first full post-training run at that scale. Weights for all three sizes are released under an open license, with hosted access also available through OpenRouter.
TL;DR
- Nex-AGI released three models under one name, mini, Pro, and Max, all sharing the same architectural lineage but scaled very differently for different deployment budgets.
- Nex-N2.5-Max runs on a 1.6-trillion-parameter Mixture-of-Experts foundation, marking the company’s first complete post-training effort at trillion-parameter scale.
- The models are built around vision-driven agent loops, meaning they don’t just generate text, they look at a screen, judge whether an action worked, and self-correct before moving on.
- Benchmark tables put Nex-N2.5-Max close to but generally behind frontier closed models like Claude Opus 5 and GPT-5.6 Sol on coding and agentic tasks, while beating them on a few specific tests like BrowseComp.
- Nex-N2.5-mini and Pro are multimodal and deployable on far smaller hardware, with Pro fitting on a single 8xH100 node and mini running on just 2xH100s.
- All three sizes are open-weight, published on Hugging Face and ModelScope, with a prebuilt Docker image using a customized sglang fork for serving.
- The release is explicitly framed as agentic infrastructure work, not just a chatbot upgrade, with a dedicated evaluation harness (NexAU) and a computer-use harness (NexCUA) built alongside the models.
Remy doesn't write the code. It manages the agents who do.
Remy runs the project. The specialists do the work. You work with the PM, not the implementers.
Why does Nex-AGI say vision matters so much for agents?
The model card’s central claim is that vision has stopped being just an input channel and has become the mechanism by which an agent checks its own work. In a long-horizon task, an agent might click a button, wait for a page to load, then need to confirm the click actually did what it was supposed to before taking the next step. That confirmation loop depends on the model interpreting a screenshot correctly and using it to plan the next action.
Nex-AGI describes this as strengthening the model’s ability “to act continuously and self-correct through visual feedback.” Practically, that means Nex-N2.5-mini and Pro are trained to operate computers and browsers directly, and to autonomously run and test code rather than just produce a code snippet and stop. This is the same underlying philosophy behind “computer use” features from other labs: instead of a model that answers questions, you get a model that drives a UI, evaluates outcomes, and retries when something breaks.
How big is Nex-N2.5-Max, and why does that matter?
Nex-N2.5-Max is built on a 1.6-trillion-parameter, text-only Mixture-of-Experts foundation model. MoE architectures activate only a subset of their total parameters for any given token, so the effective compute cost per request is much lower than a dense model of the same size would require. Nex-AGI calls this its first complete post-training effort at trillion-parameter scale, distinguishing it from the smaller mini and Pro models, which continue building on the multimodal foundations of the earlier Nex-N2 generation rather than starting from a new base model.
Unlike mini and Pro, Max is text-only. It’s positioned for scientific research, knowledge work, and complex productivity tasks rather than direct screen operation. The company frames the Max training run partly as a proving ground, calling it “valuable practical experience for training agentic capabilities in even larger models,” which suggests this scale of MoE post-training is itself a new capability the team is developing, not just a bigger checkpoint of the same recipe.
How does Nex-N2.5 perform on benchmarks?
Nex-AGI published benchmark tables comparing all three model sizes against a set of current frontier models, referred to in the model card as Claude Opus 5, GPT-5.6 Sol, Kimi-K3, GLM-5.3, DeepSeek-V4-Pro-0813, and Qwen3.8-Max.
On coding benchmarks, Nex-N2.5-Max scored 86.1 on Terminal-Bench 2.1, 65.7 on SWE-Bench Pro, and 65.6 on DeepSWE v1.1. Those numbers sit close to but below the top scores in the comparison set (Claude Opus 5 led at 89.1, 79.2, and 73.7 respectively). Nex-N2.5-Pro trailed further behind Max on every coding benchmark, and mini trailed further still, which is the expected pattern for a scaled-down model family.
On agentic and browsing tasks, the picture is more competitive. Max scored 92.6 on BrowseComp, actually topping every listed competitor. It also scored 50.2 on AutomationBench, just short of Claude Opus 5’s 50.3, and tied Kimi-K3 and Claude Opus 5 at 76.5 on Toolathlon Verified alongside them. On GDPval-AA v2 and Job Bench, Max scored 1713 and 53.6, both behind the frontier leaders but ahead of several other named competitors.
On multimodal and computer-use benchmarks, Nex-N2.5-Pro (mini and Pro are the multimodal models; Max is text-only and wasn’t evaluated here) posted 82.2 on OSWorld-Verified and 56.4 on OSWorld-2, both below the top scorers in that table but within range of models like MiniMax-M3 and GPT-5.6 Sol. Pro’s strongest relative result was OSWorld-G, a grounding benchmark, where it scored 87.4, the best result in that column ahead of every other listed model including Qwen3.8-Max’s 84.9.
Nex-AGI notes that most scores without a public source come from its own evaluations, run with temperature 0.7, top_p 0.95, and top_k 40, using an internal coding harness called NexAU and a computer-use harness called NexCUA that the company says will be open-sourced separately.
Is Nex-N2.5 worth deploying over closed frontier models?
That depends on what you’re optimizing for. On raw benchmark scores, Nex-N2.5-Max generally sits a step behind the top closed models like Claude Opus 5 and GPT-5.6 Sol across coding and several agentic categories. It’s not consistently the best performer in the comparison set. Its clearest wins are narrower: leading BrowseComp, and Pro leading OSWorld-G grounding.
Where the family may be more compelling is deployment flexibility. All three sizes are open-weight, which means you can self-host rather than depend on an API. Nex-AGI provides concrete hardware guidance: Nex-N2.5-mini runs on 2xH100 GPUs, Pro fits on a single 8xH100 node, and Max requires a multi-node setup (documented at 2 nodes of 8xH200 GPUs each, 16 GPUs total) using tensor and expert parallelism. That’s a meaningful difference from closed frontier models, where you have no visibility into what hardware you’d need because you never run the model yourself.
For teams building computer-use or browser automation agents who want an open-weight option they can inspect, fine-tune, or run on their own infrastructure, Nex-N2.5 gives a graduated set of choices from a lightweight mini model up to a trillion-parameter-class system, without requiring an account with a closed-model provider.
Frequently Asked Questions
What does Nex-N2.5-Max’s architecture actually mean for cost?
Because it’s a Mixture-of-Experts model, only a portion of the 1.6 trillion total parameters activates per token, which keeps inference compute lower than a dense model of equivalent size would need. Exact active-parameter counts weren’t specified in the model card.
Can Nex-N2.5-Max do computer use or browsing tasks?
No. Max is described as a text-only foundation model. Computer use, browsing, and visually grounded agent tasks are handled by the multimodal mini and Pro models, which build on the earlier Nex-N2 generation’s multimodal foundations.
Where can I access Nex-N2.5?
Weights for mini, Pro, and Max are published on Hugging Face and ModelScope under an open license. Hosted, API-style access to Pro and mini is also available through OpenRouter.
What hardware do you need to run these models?
Nex-AGI’s documented reference deployments use 2xH100 for mini, a single 8xH100 node for Pro, and a 2-node, 16xH200 GPU cluster for Max, served through a customized sglang fork distributed as a prebuilt Docker image.
Seven tools to build an app. Or just Remy.
Editor, preview, AI agents, deploy — all in one tab. Nothing to install.
How is Nex-N2.5 different from the earlier Nex-N2 models?
Nex-N2.5-mini and Pro extend Nex-N2’s multimodal foundations with targeted improvements to computer use, browsing, and visual grounding. Max is a new addition to the family entirely, introducing a trillion-parameter-scale MoE base model that didn’t exist in the Nex-N2 generation.
