Nex-N2.5 Mini Hands-On: Testing Next AGI's Agentic Model
Hands-on test of Nex-N2.5 Mini, Next AGI's multilingual, multimodal agentic model, deployed on dual H100 GPUs with SGLang.
What is Nex-N2.5 Mini?
Nex-N2.5 Mini is a multimodal, agentic language model from Next AGI, built on a Qwen3.5 mixture-of-experts (MoE) architecture. It’s the smaller sibling in the N2.5 family, designed for long-horizon tasks such as operating computers, browsing the web, and self-correcting using visual feedback rather than just producing a single text response. It’s released under an Apache 2.0 license on Hugging Face and ships as safetensors weights split across 16 shards, meant to be served through frameworks like SGLang on multi-GPU setups.
TL;DR
- Nex-N2.5 Mini is Next AGI’s smaller agentic model, built on a Qwen3.5 MoE backbone, and supports both text and image input for multimodal reasoning.
- The model requires two 80GB-class GPUs (H100 or equivalent) to run, deployed here on RunPod using Next AGI’s own SGLang Docker container with tensor parallelism.
- In hands-on testing it produced fast inference, around 217 tokens per second, with each of the two H100s consuming roughly 66GB of VRAM.
- On coding tasks, it generated a fairly clean self-contained HTML/canvas animation, though some visual details (particle placement, one unclear scene) showed rough edges.
- A roughly 80-language terms-of-endearment test showed the model correctly flagging languages it wasn’t confident about instead of inventing plausible-sounding but wrong translations.
- Benchmark comparisons place Nex-N2.5 Mini behind frontier models like Claude Opus 5 and Gemini 3, but its larger sibling, Nex-N2.5 Pro, competes respectably against GLM 5.3, DeepSeek V4, and GPT 5.6.
- The model is positioned as a mid-tier agentic option, not a frontier leader, but one that’s honest about its own uncertainty, which matters for agentic and multilingual use cases.
How do you deploy Nex-N2.5 Mini?
Because the model needs real GPU muscle, running it isn’t a laptop exercise. The tested setup used RunPod, a cloud GPU rental service, to provision two H100 GPUs (or any 80GB-equivalent card) on demand. The process involves selecting a GPU type and count, then choosing a template that defines the software environment the pod boots into.
Rather than using a generic prebuilt template (RunPod offers many, for ComfyUI, Ollama, vLLM, and similar tools), the deployment used a custom template built around the SGLang Docker container that Next AGI specifies in the model card. This is the detail that matters: Nex-N2.5 Mini is meant to run through SGLang specifically, and matching the serving stack to what the model publisher recommends avoids configuration headaches.
Once the template was set (with 150GB of persistent storage in this case), deploying the pod took only a couple of minutes. After connecting via SSH, the model weights downloaded quickly, and the model was launched with SGLang using tensor parallelism (TP) across both GPUs. VRAM usage settled at around 66GB per GPU, consistent with a model that needs the full two-card 80GB-class configuration the model card calls for.
How fast and capable is it in practice?
Once running, the model responded quickly. On a historical-reasoning prompt, asking it to explain the fall of the Berlin Wall using slang from a fictional book, it processed and answered at approximately 217 tokens per second on the two-H100 pod. The response held together real historical facts about the Berlin Wall while layering in invented slang terms, and the reasoning trace showed it working through the blend deliberately rather than improvising blindly.
A tougher coding prompt asked the model to build a complete, self-contained HTML file with animated canvas graphics tracking the history of the doner kebab across five eras and countries, organized into tabs. The result rendered in-browser with distinct scenes (Ottoman Empire, Turkish Republic, Berlin, a European scene, and a present-day scene), each with its own animation: rotating meat textures, flame effects, and a sliding highlight to simulate rotation. Quality was uneven. Flame and smoke effects looked convincing, including a color-shift touch in the Berlin scene, but some particle effects were misplaced and one scene’s setting wasn’t clearly identifiable. It compared unfavorably to output from stronger models like Claude Opus 5, but for a mid-tier model, the result was functional and reasonably polished.
Is Nex-N2.5 Mini good at multilingual tasks?
The multilingual test asked the model to translate a term of endearment (similar to “sweetheart”) into roughly 80 languages, checking whether it would invent plausible-sounding but fake translations for languages it didn’t actually know well. This is a common failure mode for language models under-resourced language coverage, where a model will confidently generate a wrong or fabricated term rather than admit uncertainty.
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
Nex-N2.5 Mini flagged around eight languages, including Bhojpuri, Sundanese, and Malagasy, as ones where it lacked confidence, rather than fabricating an answer. The reasoning trace showed it checking itself on gendered forms, regional variation, and source reliability before committing to a translation. A few results, in languages like Hausa or Icelandic, looked questionable on inspection, but the overall behavior of declining to guess on genuinely obscure languages is a meaningfully better failure mode than confident hallucination.
How does it handle multimodal and role-play prompts?
A separate test checked multimodal reasoning combined with a forced-choice scenario. Given an AI-generated image, the model was asked to role-play as the woman pictured and make an irreversible choice between her career (her boss) and her relationship (her boyfriend), without hedging or inventing a compromise. Nex-N2.5 Mini committed to “boyfriend,” producing prose with a plausible narrative arc, consistent tense usage, and some office-memo phrasing appropriate to the setting. It didn’t dodge the binary choice, which was the actual test: many models try to soften forced-choice prompts by suggesting the character avoid deciding at all.
Where does Nex-N2.5 Mini rank against other models?
Next AGI’s own benchmark figures, visible in the model’s published materials, show Nex-N2.5 Mini trailing frontier-tier models. Claude Opus 5 and Gemini 3 lead across most coding and agentic benchmarks, and Claude Opus 5 also leads on multimodal tasks. However, the larger Nex-N2.5 Pro variant holds its own against models like GLM 5.3, DeepSeek V4, and GPT 5.6 across several tests, landing in the middle of the pack rather than at the back.
That positions the N2.5 family, and Mini specifically, as a competitive mid-tier option rather than a frontier contender. For teams that don’t need the absolute best coding or agentic score but want a model that’s honest about uncertainty, handles multilingual queries with some self-awareness, and supports multimodal input, that middle position may be a reasonable tradeoff, especially given it’s open-weight under Apache 2.0.
Is Nex-N2.5 Mini worth running?
For self-hosting, the practical requirement is two 80GB-class GPUs (H100 or equivalent), which rules out casual local deployment. Renting GPU time through a cloud provider, using the SGLang Docker image Next AGI publishes alongside the model card, is a realistic way to test or run it without owning hardware. Inference speed in testing (217 tokens per second across two H100s) suggests the model is efficient enough for interactive agentic workflows, not just batch processing.
Whether it’s worth it depends on the use case. If the priority is frontier-level coding or agentic benchmark scores, Claude Opus 5 or Gemini 3 will outperform it. If the priority is an open-weight, self-hostable model with decent multilingual honesty and multimodal support at a mid-tier capability level, Nex-N2.5 Mini is a reasonable candidate worth benchmarking against your own workloads.
Frequently Asked Questions
What hardware does Nex-N2.5 Mini need to run?
The model card specifies two GPUs, either H100s or any 80GB-equivalent card, run with tensor parallelism through SGLang. In testing, each GPU consumed around 66GB of VRAM once the model was loaded and serving.
What architecture is Nex-N2.5 Mini built on?
It uses a Qwen3.5 mixture-of-experts (MoE) architecture and supports image-text-to-text input, making it multimodal as well as a text generation model. It’s released under an Apache 2.0 license.
How does Nex-N2.5 Mini compare to Claude Opus 5 or Gemini 3?
Remy doesn't build the plumbing. It inherits it.
Other agents wire up auth, databases, models, and integrations from scratch every time you ask them to build something.
Remy ships with all of it from MindStudio — so every cycle goes into the app you actually want.
It trails both on most coding, agentic, and multimodal benchmarks according to Next AGI’s own published comparisons. It’s positioned as a mid-tier model rather than a frontier one.
Can Nex-N2.5 Mini be run without owning GPUs?
Yes. It was deployed in testing through RunPod, a cloud GPU rental platform, using a custom template built on Next AGI’s own SGLang Docker container, with two H100 GPUs provisioned on demand.
Does Nex-N2.5 Mini hallucinate on unfamiliar languages?
In a roughly 80-language translation test, it flagged around eight languages where it lacked confidence rather than inventing translations, though a handful of its answers in other languages still looked questionable on inspection.