GMK Tech Evo X5 Pro vs NVIDIA DGX Spark: Running 150GB+ LLMs on a Mini PC

This comparison tests whether a single GMK Tech Evo X5 Pro mini PC, equipped with 192GB of unified memory, can replace two linked NVIDIA DGX Sparks for running large local LLMs and coding agents. It walks the box from stock Windows through BIOS and Linux tuning to a setup where big models actually fit and run.

■ Hardware and Setup
- Specs, ports, memory-per-dollar comparison, and CPU benchmarks versus the Spark
- Stock Windows out of the box: the 128GB GPU share, Vulkan benchmarks, and dense vs. mixture-of-experts speed differences
- Why WSL cannot see the GPU on this chip, plus dual-boot Linux testing
- BIOS UMA frame buffer settings, raising the GPU share to 160GB, and Windows paging file adjustments

■ Model and Agent Testing
- Loading 142-157GB models such as GLM 5.3, DeepSeek V4 Flash, and Qwen 235B, plus quantization and context length effects
- Multi-model and agent workloads, Docker container memory usage, and how the RAM split limits concurrent agents
- Single-user, multi-user, and team-usage scenarios

■ Software Optimization
- Linux versus Windows, a one-line ComfyUI video generation fix, speculative decoding, Flash Attention, Vulkan vs. ROCm backends
- vLLM comparisons and the Halogen engine's prompt processing results

■ Media Generation and Conclusions
- Image and video generation tests with Stable Diffusion XL, Flux, and LTX, and recommendations for different use cases

This is aimed at developers and AI enthusiasts deciding between a large-memory mini PC and a DGX Spark for local inference. It covers what each machine handles well, which configurations suit which workloads, and the practical steps needed to get there.