開源模型
Nex-N2.5 Opens Three Agent Models, With the Vision and Trillion-Parameter Text Versions Showing Different Capability Boundaries
Nex AGI has released the 35B mini, multimodal Pro, and 1.6T MoE Max models, targeting browser and desktop operation as well as long-running tool use. Official results show Max approaching leading closed models in automation scores, but lagging significantly on the vision-dependent OSWorld-2 benchmark, while local deployment requires as many as 16 H200 GPUs.
Nex AGI released the weights for the Nex-N2.5 series on September 8, offering three variants: mini, Pro, and Max. mini and Pro continue the multimodal approach, using screen images to verify the results of actions and targeting web browsing, GUI interaction, and code execution. Max, by contrast, is a text-only, 1.6-trillion-parameter MoE model. All three are licensed under Apache-2.0, provide a 262,144-token context window, and use the OpenAI Chat Completions-compatible `reasoning_effort` parameter to switch among direct-answer, adaptive-reasoning, and forced-reasoning modes. The [official repository](https://github.com/nex-agi/Nex-N2.5) also specifies the Qwen3 Coder tool parser, while Max requires the DeepSeek-R1 parser to separate its reasoning traces.
The published results show that “agent capability” cannot be judged by model size alone. Max scores 50.2 on AutomationBench v1.0.6, just 0.1 points below the officially listed Claude Opus 5 result, and achieves 86.1 on Terminal-Bench 2.1. However, on OSWorld-2, which requires understanding on-screen content and operating a computer, the text-only Max scores just 30.5, compared with 56.4 for the multimodal Pro and 68.3 for Claude Opus 5. This contrast suggests that scaling up a text model is not enough to replace visual grounding and action-feedback loops.
Deployment cost is equally important. The [mini model card](https://huggingface.co/nex-agi/Nex-N2.5-mini) recommends two H100 GPUs, while Pro requires eight H100 GPUs. Max’s reference configuration uses two nodes with a total of 16 H200 GPUs and depends on Nex’s modified SGLang image, FP8 KV cache, DeepEP, and DeepGEMM. Engineering teams should next watch for NexCUA, the company’s evaluation framework that has not yet been open-sourced, as well as third-party reproductions and each model’s performance in failure recovery, permission control, and the cost of long-running tasks. For now, the scores come primarily from the vendor and should not be treated as independently verified.