Fix Outdated LM Studio Headless Runtime on NVIDIA DGX Spark

Published 09/08/2026

Description

Running LM Studio headless on an NVIDIA DGX Spark or Ubuntu server is convenient: the local LLM service can start automatically after reboot without logging in to a graphical desktop. Contents: 0:00 - Introduction 0:14 - Description 1:30 - The Problem 3:12 - The Solution 4:38 - Conclusion But there is a maintenance issue to watch for. LM Studio and the headless llmster service use inference runtimes based on llama.cpp. Updating LM Studio on a desktop client does not necessarily mean that the runtime used by your remote headless service is current. When a newer model requires a newer runtime, the model may fail to load even though the LMS service is online and reachable. In this video, I show how I update the headless LM Studio stack on my NVIDIA DGX Spark / ASUS GX10 so newer models can load again through a remote LM Studio client. LM Studio documents llmster as its standalone daemon for GUI-less deployments and recommends it for headless/server-style use. The lms CLI can manage both the daemon and inference runtimes. What this video covers Running LM Studio headlessly with llmster Checking LMS service status on Ubuntu and DGX Spark Connecting a Windows LM Studio client to a remote DGX Spark Diagnosing a model load failure caused by an outdated runtime Comparing the runtime available in LM Studio with the headless runtime Stopping the LM Studio server and daemon safely Updating the llmster daemon Updating the llama.cpp runtime Listing installed runtimes Selecting the required CUDA 13 Linux ARM runtime Restarting the headless service Verifying that a previously failing local LLM loads remotely LM Studio provides lms daemon update to update llmster and lms runtime update llama.cpp to fetch the latest llama.cpp runtime. Its documentation also notes that the daemon should be stopped before updating and started again afterward. If this helped, please like the video, subscribe, and share what hardware, Ubuntu image, m