Nex-N2.5-mini

// Overview

Nex-N2.5-mini is the entry tier of the Nex-N2.5 family Nex AGI published on 8 September 2026, and the cheapest way to run one of these agent models on hardware you control. The Hugging Face weight index totals 35,107,181,936 parameters in BF16 across 16 safetensors shards, and the config declares a `Qwen3_5MoeForConditionalGeneration` architecture — a multimodal mixture-of-experts stack built on a Qwen 3.5 base.

Size is the point. Nex AGI documents mini running on a single node with just two H100 GPUs, against eight for Nex-N2.5-Pro and two full nodes of sixteen H200s for Nex-N2.5-Max. It keeps the same 262,144-token context as the larger tiers, so the saving is in GPUs rather than in what the model can hold.

Like Pro, mini is multimodal and shares the family's central idea: vision is a feedback channel, not just an input. The model looks at a screen, a browser or a program it launched, checks the result against the goal, and revises. On OSWorld-G, the grounding benchmark, mini scores 82.9 — behind Pro's 87.4 but ahead of Claude Opus 5 on 76.8 and GPT-5.6 Sol on 77.7 in Nex AGI's published table.

The trade-off is visible on hard coding and knowledge work, where mini drops well behind its siblings: 43.8 on SWE-Bench Pro against 65.7 for Max, and 28.5 on Job Bench against 53.6. It is a model to spend on the many small steps of an agent run rather than on the one hard problem in the middle of it.

// Benchmarks

This model's scores

1. OmniDoc89.7%
2. BrowseComp83.4%
3. OSWorld-G82.9%
4. Terminal-Bench 2.173.4%
5. OSWorld-Verified71.2%
6. WebArena-Verified63.4%
7. Toolathlon Verified54.6%
8. SWE-Bench Pro43.8%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Strengths

• Serves on a single node with two H100s, the smallest footprint in the Nex-N2.5 family
• Full 262,144-token context, the same window as the 1.6T Max tier
• Multimodal grounding that holds up: 82.9 on OSWorld-G, ahead of Claude Opus 5 (76.8) in the published table
• Apache-2.0 weights published on Hugging Face, with GGUF-free direct BF16 shards and a documented SGLang launch command
• Free hosted access on OpenRouter as nex-agi/nex-n2.5-mini:free, so it can be tested before any GPU spend

// Best for

• Sub-agent and worker roles inside a larger agent system, where cost per call matters more than peak ability
• GUI and browser automation on a budget, using the grounding scores rather than a frontier model
• Self-hosted agent prototypes on a single two-GPU box
• Long-context document and screen work that would be expensive to send to a hosted frontier API

// Nex-N2.5 — every version

The full lineage of the Nex-N2.5 line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.