Summary
Fahd Mirza provides a hands-on deployment walkthrough and capability test of Nex AGI’s N2.5 Mini, the smaller model in the N2.5 agentic lineup designed for long-horizon tasks including computer operation, web browsing, and visual self-correction. The setup section is detailed: Mirza provisions two H100 GPUs (80 GB each) on RunPod using the model’s own SGLang Docker container, creates a custom template, and downloads the model — demonstrating a full self-hosting workflow from scratch for viewers unfamiliar with cloud GPU deployments.
Once running, the model achieves 217 tokens per second and consumes approximately 66 GB of VRAM per GPU. Mirza runs a series of tests: a creative reasoning prompt combining Berlin Wall history with fictional slang, a complex single-file HTML canvas simulation of döner kebab’s historical evolution across five eras, and agentic tool-calling tasks. Results are generally in line with the model’s benchmark positioning — strong on logic, math, and structured code generation, weaker on long-form prose and complex visual rendering.
Mirza positions N2.5 Mini honestly against published benchmarks: it trails frontier leaders like Claude Opus 5 and Gemini K3 on most coding and multimodal tasks, but holds its own against mid-tier models such as GLM 5.3, DeepSeek V4, and GPT 5.6, landing in the middle of a competitive pack. The video is most useful for practitioners evaluating whether N2.5 Mini is worth the two-GPU minimum and how it fits into an agentic pipeline alongside larger orchestrator models.
📺 Source: Fahd Mirza · Published September 10, 2026
🏷️ Format: Review
