A first local-AI setup should prove one useful workflow, not create a collection of software and model files.
The simplest beginner route in Private AI on Your PC uses LM Studio as the worked example because it provides a graphical way to find, download, load, and chat with supported local models while showing information about model variants and device fit.
The exact interface can change, so follow the workflow rather than memorizing a button position.
Before you install anything
First complete the PC hardware-fit check.
Record:
- system RAM;
- GPU and dedicated VRAM, if present;
- free storage;
- normal background memory use.
The book’s September 2026 source check records LM Studio guidance recommending at least 16 GB RAM and at least 4 GB dedicated VRAM for its Windows application, with AVX2 required on x64. Smaller models can still be useful on some systems without a dedicated GPU, but application requirements and model behaviour change, so confirm the vendor’s current documentation before relying on a specific requirement.
Step 1: install from the official source
Use the official LM Studio website rather than a third-party download mirror.
If the computer has a small system drive and a larger secondary drive, decide where model files should be stored before downloading several gigabytes of data.
Step 2: choose one mainstream instruct/chat model
Do not start with the largest model in the catalogue.
Look for a general-purpose instruct or chat model and use the application’s device-fit information.
If several quantisations are offered, a 4-bit or 5-bit option that fits comfortably is a practical starting point.
The objective is not to prove that the biggest file can load. The objective is to keep Windows responsive while the model performs a useful task.
Step 3: download one model completely
Model downloads can be several gigabytes.
Let the download finish before testing. Once the model is in the local library, load it into a chat session.
Keep Task Manager open while the model loads.
If available memory collapses and the whole PC becomes sluggish, unload the model and step down in size before changing advanced settings.
Step 4: run a three-prompt smoke test
Use simple prompts that test basic instruction following.
For example:
- Draft a short professional email declining a meeting and offering two alternatives.
- Summarize a paragraph you wrote yourself into five factual bullets.
- Explain RAM versus storage to a non-technical adult.
You are not testing whether the local model is smarter than a leading cloud system.
You are checking whether it is responsive, follows format instructions, and preserves supplied facts closely enough for practical use.
Step 5: run one factual-preservation test
Give the model a short source block that contains names, dates, and numbers.
Ask it to extract those values without changing them.
If the answer is fluent but alters factual details, do not make that model your default for factual work.
A useful model needs both acceptable quality and acceptable reliability for the tasks you intend to give it.
Step 6: test the local session offline
After the model and required runtime are downloaded:
- run a normal prompt while connected;
- close the application;
- temporarily disconnect Wi-Fi or Ethernet;
- reopen the application;
- select the already-downloaded local model;
- run the same prompt.
If document chat is part of your intended workflow, attach a small local text document and ask a source-only question while still disconnected.
A successful test is evidence that the tested workflow can run without a live network connection.
It is not proof that the whole application makes zero network requests when connected, and it says nothing about other software on the computer.
Step 7: confirm the active model is actually local
A good answer does not prove that the answer was generated locally.
Applications can add optional hosted models or online features over time.
Before using private material, confirm that:
- the selected model is the downloaded local model;
- web search is not enabled unless you intend to use it;
- cloud or hosted model options are not active for that session;
- you understand where the application stores chats and local data.
Use the privacy-boundary guide before treating the setup as private.
Step 8: record the baseline
Write down:
- application version;
- model name;
- quantisation or file variant;
- context setting;
- a rough speed impression;
- whether the PC remained responsive;
- whether the factual test passed;
- whether the offline test passed.
This baseline becomes useful when a later update, larger model, driver change, or context increase makes the system behave differently.
Keep the first setup deliberately narrow
The durable sequence is:
download → load → test → verify → reuse
Do not install three runtimes and five versions of the same model on day one.
Prove one setup for one task. Save the prompt that worked. Use it for several days. Add another tool only when you can name the feature you actually need.
If the model is slow or unstable, follow the local AI troubleshooting order before reinstalling software or buying hardware.