Nex-AGI releases Nex-N2.5-mini, a 35B-parameter agentic model

The model nex-agi/Nex-N2.5-mini has been published on HuggingFace with the text-generation task. It is the smaller model in the new Nex-N2.5 family, designed for long-horizon agentic tasks in real-world environments: computer control, browser control, autonomous program execution and verification. The author notes that vision here is not just an input modality but an interface through which the agent checks results and advances the task.

What it means

The model has 35.1 billion parameters and uses a MoE architecture. The full weights repository takes up 65.4 GB, with the largest file at 4.6 GB. The license is apache-2.0, and it runs via transformers. There are already 8 quantized builds available on the platform: GGUF, MLX, NVFP4, FP8. The lightest working build is 17.3 GB, which is julianmb/Nex-N2.5-mini-ROCmFP4-GGUF. Running the full model locally requires a substantial amount of VRAM, but quantized variants lower the entry barrier. ROCm and GGUF builds make it possible to try the model on your own hardware without loading the full 65 GB. For more news on local AI tools, see the CPU3D news feed.