Nex-N2.5-mini

Nex-N2.5-mini is the entry model of the Nex-N2.5 family, published with open weights on 8 September 2026. Like the Pro, it is not a model Nex AGI trained from scratch: the company's website names Qwen3.5-35B-A3B-Base as its foundation, while the Hugging Face card mentions only the multimodal groundwork of the earlier Nex-N2. What Nex AGI adds is post-training for agent work — operating a computer, browsing, running and testing code with visual feedback. As a small model it loses most comparisons, and the maker publishes those losses rather than hiding them: 73.4 on Terminal-Bench 2.1 against 82.7 for the Pro, 43.8 on SWE-Bench Pro against 61.2, 28.5 on Job Bench against 41.4. One result breaks the pattern and is the reason this profile exists: on OSWorld-G, which measures whether a model can point at the right element on a screen, it scores 82.9 — ahead of Claude Opus 5 at 76.8, GPT-5.6 Sol at 77.7 and Kimi-K3 at 79.6. Grounding a click, unlike reasoning about a task, turns out not to need a big model; it needs training on screens. The architecture is the Qwen3.5 sparse mixture of experts in its small configuration: 40 layers, 256 routed experts with 8 activated per token, hidden size 4096, vocabulary 248,320, a hybrid linear-attention path and a declared window of 262,144 tokens. Weights are published FP8 block-quantised and the maker's recipe runs the model on two H100 cards, which puts it within reach of a single workstation-class server. The model is hosted free of charge on OpenRouter, with no paid listing at launch, and community quantisations — GGUF, MLX, NVFP4 — appeared within days, several of them with more downloads than the original weights. Thinking is adaptive by default, function calling uses the Qwen tool format, and all benchmark numbers on this page come from the maker.

▸Videos

No videos yet.

▸Reviews