The simple part is real. The caveats about hardware, model size, offline use and privacy are where a smooth demo can turn into a frustrating evening. This guide walks through the three steps and adds those caveats.
What the walkthrough covers
Hunt's setup uses LM Studio, a free desktop app for downloading, managing and chatting with local models. He had used Ollama, a similar tool, on a desktop before. For this test he chose LM Studio for its clean interface.
The sequence is:
- Install LM Studio.
- Download a model inside the app.
- Load the model and start chatting.
LM Studio's own documentation describes the same flow. It is a cross-platform desktop application that runs large language models locally. Its docs say the app handles model discovery, downloading and local inference.
Step 1: Install LM Studio
Download the installer from the official LM Studio site and pick the Windows build that matches your PC. LM Studio's documentation says it runs on Windows (x64 and ARM), and Hunt notes it also runs on Linux and macOS.
Before you install, check the basics. Third-party guides, not the vendor's own summary, put the Windows floor like this:
- AVX2 instruction set support is required on x64 CPUs. An older processor without AVX2 is not merely slow. It is incompatible.
- ARM devices built on the Snapdragon X Elite are also supported.
- 16GB of RAM is recommended, along with at least 4GB of dedicated VRAM for discrete graphics.
Treat those numbers as a minimum for launching the app. They don't tell you whether a given model will run well.
Step 2: Pick a model that fits your machine
This is the step where people go wrong. Hunt's test model was a 27-billion-parameter model that he calls Qwen3.8. I couldn't independently confirm that name or listing, so treat it as his description. He is clear that it is too heavy for the average PC. He chose it to stress the RTX Spark laptop and its large unified memory pool.
For everyone else, his advice is to start small. He suggests a 7B or 8B model for beginners. His rules of thumb are:
| Model size | Rough GPU memory target (Hunt's guidance) |
|---|---|
| 3B–8B parameters | At least 8GB of VRAM |
| 8B–14B parameters | About 12GB of VRAM |
| 30B and up | Workstation-class VRAM or a unified memory pool |
These are one reviewer's guidelines, not LM Studio specifications. Other guides run lower. One third-party estimate says 8–12GB of VRAM supports 7–14B models at 4- or 5-bit quantization. Quantization is how much a model's weights are compressed, and it changes memory needs a lot. Context length matters too. A useful way to think about it is that model file size, plus context cache, plus operating system overhead, must be smaller than the fast memory you have.
LM Studio's built-in model search helps here. Hunt says it flags compatibility and performance, and he suggests looking for the "Full GPU Offload Possible" tag. That tag is a good sign, but it isn't a promise about speed in every workload.
Once you've picked a model, click Download and let it finish. These files are big. One guide says a model weighs from 2GB to more than 60GB, so check your free disk space first.
Step 3: Load the model and chat
LM Studio's documentation lays out the loading step. Go to the Chat tab, open the model loader and select a model you've downloaded. You can optionally set load parameters. It also explains what "loading" means: the model has to be loaded entirely into memory, either VRAM or system RAM, to function properly. That is why a model that is too large will load slowly, fail, or spill onto slower memory.
Hunt's route is to click Chat in the top-left corner, then "Select a model to load" at the top center. Pick your model, wait a few moments, and start typing.
What "private" and "offline" really mean
This is where the headline claim needs care.
- Offline after setup. You need a connection to download the app and the model. After that, LM Studio's docs say it can operate entirely offline, as long as you get some model files first. I'm paraphrasing the docs here. Hunt's FAQ says the same.
- Private inference. Running the model on your PC means your prompts don't have to go to a cloud AI provider. That is a narrower claim than "your whole computer is private." It says nothing about other apps, browser activity or optional integrations.
- Network serving. LM Studio can also serve models over an API. Its docs say it can serve local models on OpenAI-like endpoints, locally and on the network. If you turn that on, check whether it is reachable from other devices. Leave it local unless you mean to share it.
Is it really free?
The app and many open-weight models cost nothing to download. You still pay in other ways:
- Hardware capable of running the model.
- Storage for large model files.
- Battery life and power. Heavy use drains a laptop battery quickly.
- Model licenses, which differ. "Open" doesn't mean every model has the same terms.
Hunt also concedes the trade-off. Local models are usually weaker than the top cloud models. That is fine for chat, summaries and light study. Heavier users will probably still want a paid cloud subscription for some work.
Troubleshooting common problems
- Responses are very slow. Try a smaller or more heavily quantized model, or reduce the context length. One guide also advises making sure GPU offloading is enabled if you have a dedicated card.
- The model won't load or crashes. It probably doesn't fit in memory. Pick a smaller variant.
- Nothing runs on an older PC. Check for AVX2 support on x64 machines.
- CPU-only. It works, but Hunt doesn't recommend it, and he expects far slower performance.
NPUs and Copilot
Hunt's FAQ answers two common questions:
- Can local LLMs run on a laptop NPU? He says newer hardware, such as the Snapdragon X2 line at 80 TOPS, can run lighter local models. First-generation NPUs are weaker.
- Is Copilot local AI? He says the Copilot app on Windows 11 relies on the cloud and won't work without a connection. He says Copilot+ features like Recall and Studio Effects use the NPU for local acceleration instead.
Don't read the RTX Spark as a requirement
The test machine is a premium laptop. NVIDIA announced RTX Spark PCs on October 16, 2026, with unified memory shared between CPU and GPU. I couldn't open NVIDIA's announcement directly, so the platform figures come from the research brief only. Those figures are maximums, not what every configuration ships with. The shortened "Pro 9n" name in the article may also refer to the Yoga Pro 9n, which I couldn't confirm.
The practical takeaway is that you don't need this hardware to try local AI. A gaming PC with a decent GPU and a small model is a better starting point for most people. It is cheaper, and you can find out whether you'll actually use it.
The bottom line
The three-step install is as easy as Hunt says. The harder part is choosing a model your PC can hold. Start with a 7B or 8B model, watch your memory, and think of "private" as meaning your prompts stay on the machine. Don't expect it to match the biggest cloud models.