Nex-N2.5 arrives open source: 0.1 behind Opus 5 on one benchmark, less than half on computer use
Shanghai-based Nex AGI released three Apache-2.0 agent models. The announcement leads with a 0.1-point gap on a single benchmark; the model card's own table shows 30.5 against 68.3 on OSWorld-2.
Shanghai-based Nex AGI has released the Nex-N2.5 family as open source. Three models: mini (35 billion parameters), Pro, and Max, a 1.6-trillion-parameter text-only mixture-of-experts model. All three carry an Apache-2.0 licence and a 262,144-token context window. The organisation describes itself as "an innovation alliance initiated by the Shanghai Innovation Institute".
The number the announcement leads with: Max scores 50.2 on AutomationBench v1.0.6, just 0.1 behind Claude Opus 5 at 50.3. That figure is accurate. The rest of the table on the same model card does not appear in the announcement.
| Benchmark | N2.5-Max | N2.5-Pro | Claude Opus 5 | Qwen3.8-Max |
|---|---|---|---|---|
| AutomationBench v1.0.6 | 50.2 | 44.2 | 50.3 | 39.8 |
| OSWorld-2 | 30.5 | 56.4 | 68.3 | 46.7 |
| Terminal-Bench 2.1 | 86.1 | 82.7 | 89.1 | 86.6 |
| SWE-Bench Pro | 65.7 | 61.2 | 79.2 | 67.7 |
The card tells a different story than the announcement
Start with OSWorld-2. Max scores 30.5 there; Opus 5 scores 68.3. The "0.1 point" closeness belongs to one benchmark only, and on the test that measures driving a computer the gap is more than twofold.
The second row is stranger: the smaller Pro beats the larger Max by 25.9 points on the same benchmark, 56.4 against 30.5. The model card offers no explanation. Pro is multimodal and Max is text-only, and OSWorld-2 measures working from what is on screen, so that may be where the gap comes from. That is our inference; the card does not say it.
On none of the published rows does a Nex-N2.5 model beat Claude Opus 5. Pro does beat Qwen3.8-Max on some rows (AutomationBench 44.2 against 39.8, OSWorld-2 56.4 against 46.7) and trails it on others (Terminal-Bench 82.7 against 86.6, SWE-Bench Pro 61.2 against 67.7, OSWorld-Verified 82.2 against 86.1).
Open source, but running it is another matter
For Max the model card recommends two nodes and 16 H200 GPUs, and a fork of SGLang patched by Nex. The Apache-2.0 licence comes with the weights, but the number of people who can stand up that hardware is small. Open source here means downloadable, not runnable by everyone.
The capabilities the card claims: continuous action, self-correction through visual feedback, working across Blender and CAD tools, and running programs independently in a browser and on a desktop. Every figure in this article is self-reported by the vendor; no independent evaluation has been published.
Sources: Nex-N2.5-Max model card, Nex-N2.5-Pro model card, Nex AGI.
