[Local LLM 06] Move Model Storage and Check Offline Operation

Building a Local LLM · 06/30 · Basics

Full series and learning sequence · Practice code ZIP

Several downloaded models can consume a disk sooner than expected. Identify the model storage directory separately from the Ollama program installation. Knowing which files belong in each location makes a disk move or a later rebuild easier to plan.

Prepare storage space to move Ollama and the model. Before changing the path, write down the current model list and storage location.

Saved and loaded models

ollama list
ollama ps
ollama stop qwen3:4b-instruct

When I run stop, the model is taken down from memory. Use ollama rm 모델이름 to clean up disk space. Deleted models may need to be downloaded again. Also, because multiple models share the same blob, the sum of the listed sizes does not necessarily equal the actual disk footprint.

If you pick and delete large files from the blobs folder, the files pointed to by the manifest may disappear. Model cleanup is done using CLI. It is safer to look at the files in the folder when checking the storage location or backing up.

Change storage location to D drive

The example below is when drive D exists and is writable. Change the path if you use a different drive. This step specifies a save location and does not automatically move existing models.

New-Item -ItemType Directory -Force D:\AI\OllamaModels
[Environment]::SetEnvironmentVariable("OLLAMA_MODELS", "D:\AI\OllamaModels", "User")

Completely quit Ollama from the tray and relaunch it from the Start menu. If the list appears empty in the new PowerShell, check that it is reading a new repository. The simplest way is to pull back only the models you need to a new location. After checking the download time and capacity, decide whether to clean up the old storage.

If you have many existing models, you can also close the app and copy the entire save folder. It does not select and move only some blobs. Apply the new path, keeping the original path and folder, and clean up the old copy once both the list and actual answers are verified.

Disconnect the network and check the inference

Disconnect from the network with the model you have downloaded and try running the same text questions again. Options with :cloud in the name are excluded from this offline verification. Web browsing, external embedding, remote tools, and checking for updates can all use the network independently of model inference.

ollama run qwen3:4b-instruct "물의 끓는점을 설명하되 고도에 따른 차이도 짧게 알려 줘."

What this confirmation tells us is that the request was processed without an external inference server. This should not be interpreted broadly to mean that other features of the app do not communicate at all times. In environments where export control is required, firewall policies and connection records are also reviewed.

Archive files and execution conditions

ollama --version | Out-File -Encoding utf8 .\ollama-version.txt
ollama list | Out-File -Encoding utf8 .\models.txt
ollama show qwen3:4b-instruct | Out-File -Encoding utf8 .\model-info.txt

The same model tag can be updated with different content later. Leave your context settings and questions along with model ID, size, and download date. When passing on files to others, also check the redistribution conditions for the model.

When recovering, in addition to the model file, system prompts, RAG text, index creation conditions, and evaluation table are also required. If the size is a burden, you can download the model again and back up small settings and the original text first. In the final article, we will compile this item into operational records.

reference material

• Set Windows model path
• CLI Model Management
• Ollama FAQ

Documentation and model tags checked: October 18, 2026. Primary lab environment: Windows 11 x64.

---