Set Up DGX Spark with Qwen 3.6 27B Dense

This article details the complete process of deploying the Qwen 3.6 27B BF16 model on an NVIDIA DGX Spark. The hardware utilizes the Grace-Blackwell architecture with 128GB of unified memory. By disabling the desktop environment, sleep mechanisms, and hardening SSH security configurations, the setup ensures maximum resource utilization and system stability. The deployment architecture uses vLLM to provide an OpenAI-compatible API, with OllamaGateway handling authentication and model alias management. Key configurations include full BF16 precision, 95% memory utilization, a 256K context window, and prefix caching, balancing quality and performance. Tests show that, limited by memory bandwidth, inference speed is approximately 5.5-16 tok/s, but it supports switching thinking modes and tool calls per request. The system uses systemd for automatic startup and fault recovery, with a cold start time of about 150 seconds. The overall solution involves no quantization and no external dependencies, with sufficient memory headroom, making it suitable for high-reliability local LLM service deployment within a local area network.