How to Run DeepSeek or Llama Locally on Windows With LM Studio

Updated October 2026. Checked against LM Studio 0.4.26, the LM Studio docs, and the DeepSeek and Meta model cards on Hugging Face on 24 October 2026.

Short answer: install LM Studio from lmstudio.ai/download, open the Discover tab, search for DeepSeek R1 0528 Qwen3 8B or Llama 3.1 8B, download the Q4_K_M file (about 5 GB), then load it in the Chat tab and start typing. Both the app and the models are free. On a Windows PC with 16 GB of RAM and a graphics card with 6 to 8 GB of VRAM, that's a comfortable setup. With less, pick a smaller model.

One thing trips up almost everyone, so I'll say it upfront. The "real" DeepSeek you use in the DeepSeek app is not something a home PC can run. What you run locally is a distilled DeepSeek model: a small model trained on DeepSeek's reasoning. It's still good, and it's fully offline. Here's how to set it up properly.

Jump to:

• What you need
• Which models fit on a PC
• Step 1: Check RAM, VRAM and CPU
• Step 2: Install LM Studio
• Step 3: Download a model
• Step 4: Load it and set GPU offload
• Step 5: Chat and tune settings
• Step 6: Local API (optional)
• If it's slow or won't load
• FAQ

What you need before you start

LM Studio's system requirements page is short, and these are the bits that matter on Windows:

• A Windows PC on x64 or ARM. Snapdragon X Elite laptops are supported, not just Intel and AMD.
• A CPU with AVX2 (on x64). This is a hard requirement, not a suggestion.
• 16 GB of RAM recommended. You can get by with 8 GB if you stick to small models.
• 4 GB of dedicated VRAM or more recommended. A GPU isn't strictly needed, but it's the difference between "usable" and "go make tea".
• Free disk space. Each model is a single file of roughly 1 to 20 GB for the sizes in this guide, so 20 GB free is a sensible floor.

That's it. No Python, no command line, no account. LM Studio is free under its terms of use for personal use and for internal use at work, so the cost is $0 (₹0) in both the US and India.

Which DeepSeek and Llama models actually run on a PC

Models come as GGUF files, the format used by llama.cpp, which is the engine LM Studio runs on Windows. Each model is offered at several "quantization" levels, labelled things like Q3_K_L, Q4_K_M or Q8_0. Lower numbers mean a smaller, more compressed file and slightly worse answers. LM Studio's own advice is to pick a 4-bit option or higher if your machine can handle it, and Q4_K_M is the one I default to.

I pulled these file sizes from the lmstudio-community GGUF repos on Hugging Face today. They're the Q4_K_M downloads unless noted:

Two rows at the bottom are there to save you a wasted afternoon. LM Studio's catalog page for DeepSeek V4 Flash lists a minimum of 156 GB of system memory. DeepSeek's newest release, V4.1 Flash, is a 552B-parameter model according to its model card. Llama 4 Scout is smaller but still needs well over 64 GB. None of that is happening on a gaming laptop.

So the honest version of "run DeepSeek locally" in 2026 is the R1 distills. My pick is DeepSeek-R1-0528-Qwen3-8B. DeepSeek built it by training Qwen3 8B on reasoning from its bigger R1-0528 model, and its model card reports it beating the plain Qwen3 8B on the AIME 2024 maths benchmark by 10 points. LM Studio lists its minimum system memory as 4 GB. For Llama, Llama 3.1 8B Instruct is the safe all-rounder, and Llama 3.2 3B is the one to grab if your PC is short on memory. Meta lists Hindi among Llama 3.2's officially supported languages, which is handy if you're in India.

Step 1: Check your RAM, VRAM and CPU

Before downloading anything, find out what you're working with. It takes a minute.

1. Press Ctrl + Shift + Esc to open Task Manager and click Performance.
2. Click Memory. The total at the top right is your RAM.
3. Click GPU 0 (and GPU 1 if you have it). Look for Dedicated GPU memory. That's your VRAM. Integrated graphics will show only shared memory, which is just system RAM.
4. Click CPU and note the model name shown at the top.

For AVX2, search your CPU's name on Intel's or AMD's spec pages. Nearly every Intel Core and AMD Ryzen chip of recent years has it. Some older budget Pentium and Celeron chips don't, and on those LM Studio simply won't work.

Then match the numbers to the table. My rule of thumb: if the model file fits inside your VRAM with a gigabyte or two to spare, it'll run entirely on the GPU and feel quick. If it doesn't, LM Studio can split it between GPU and RAM. That works, it's just slower. An 8 GB card like an RTX 4060 handles the 7B and 8B models fine. A 12 GB card opens up the 14B.

Step 2: Install LM Studio

1. Go to lmstudio.ai/download. The current Windows version is 0.4.26.
2. Download the Windows installer and run it. It installs like any other app.
3. Open LM Studio. You'll land on an empty chat, because there's no model yet.

Don't get confused by the download page. It now leads with Bionic, a separate agent app from the same company. You don't need Bionic for any of this. Scroll down to the LM Studio button.

The install also gives you lms, LM Studio's command-line tool. You'll only need it if you like terminals, but it's useful later. LM Studio has to be opened once before lms works.

Step 3: Download a DeepSeek or Llama model

1. Open the Discover tab, or press Ctrl + 2 from anywhere in the app.
2. Search for deepseek r1 or llama. You can also paste a full Hugging Face URL into the search box.
3. Open the model you want, for example DeepSeek R1 0528 Qwen3 8B.
4. Pick the download option. Choose Q4_K_M if you're unsure, then click Download.

Watch the size before you click. Five gigabytes on a slow connection can take a while, and if your C: drive is already tight, the download will fill it fast. If Windows feels sluggish while you're downloading, that's the disk working, and my guide to 100% disk usage on Windows 11 explains what to check if it stays that way.

Prefer the terminal? This works too, and the docs show the same pattern for picking a quantization:

lms get llama-3.1-8b@q4_k_m

Models are stored in LM Studio's models folder. If you want them on a bigger drive, change the location from the My Models tab (Ctrl + 3). If you already have a GGUF file from somewhere else, lms import path\to\model.gguf moves it in without downloading again.

Step 4: Load the model and set GPU offload

"Loading" just means LM Studio copies the model into memory so it can answer. Here's how:

1. Open the Chat tab.
2. Click the model picker at the top and choose the model you downloaded.
3. Before it loads, you'll see load settings. The two that matter are GPU Offload and Context Length.
4. Click Load Model.

GPU Offload decides how much of the model sits on your graphics card. Slide it to max if the model fits in your VRAM. If loading fails or your PC grinds, bring it down a few notches and try again. With no dedicated GPU, set it to zero and let the CPU do the work.

Context Length is how much text the model can keep in mind at once. Bigger sounds better, but it eats memory on top of the model file. For normal chatting, I leave the default alone. Raise it only when you're pasting long documents.

If you'd like a number instead of trial and error, the CLI has an estimator. It prints the GPU and total memory a model will need, without loading it:

lms load --estimate-only <model-key> --context-length 8192

Run lms ls to see the exact model keys on your machine.

Step 5: Chat and tune the settings

Type a question and hit Enter. With a DeepSeek R1 model, you'll notice it "thinks" first, writing out a chain of reasoning before the answer. That's normal and it's the whole point of these models. It also means answers take longer than with Llama, so don't judge speed on the first reply.

DeepSeek gives some specific advice for the R1 series on its model card, and it's worth following:

• Set temperature between 0.5 and 0.7. DeepSeek recommends 0.6. Too high and it rambles or loops.
• For the original R1 distills (Qwen 1.5B to 32B and Llama 8B), skip the system prompt and put every instruction in your message instead.
• For maths, ask it to "reason step by step" and put the final answer at the end.

The newer R1-0528 Qwen3 8B is less fussy. Its model card says system prompts are now supported, so you can give it a standing instruction like "answer in Hinglish" or "keep it short".

Llama 3.1 and 3.2 don't think out loud, so they reply faster and feel more like a normal chatbot. I use Llama for drafting and summaries, and switch to the DeepSeek distill for anything with logic or numbers in it. Switching is two clicks in the model picker, so try both on the same question.

A practical tip for laptops: plug in. Local models push the CPU and GPU hard, and most laptops hold back performance on battery. If you've noticed your battery draining faster than it used to, check your laptop battery health before blaming the model.

Step 6: Use it as a local API (optional)

This is the part that makes LM Studio more than a chat window. It can act as an OpenAI-compatible server on your own PC, so coding tools and scripts that expect the OpenAI API can talk to your local model instead.

1. If you can't see the Developer tab, turn on Developer Mode in Settings > Developer.
2. Go to the Developer tab and toggle Start server. Or run lms server start.
3. Point your app at http://localhost:1234/v1. Port 1234 is the default.

A quick test from Command Prompt (Windows ships with curl), using the model key from lms ls:

curl http://localhost:1234/v1/chat/completions -H "Content-Type: application/json" -d "{\"model\":\"your-model-key\",\"messages\":[{\"role\":\"user\",\"content\":\"Say hello\"}]}"

By default the server only listens on your own machine. There's a Serve on Local Network option if you want your phone or another PC to use it, but LM Studio itself warns that this exposes the server beyond localhost and recommends turning on authentication. If you do it, keep it on your home network and don't forward the port on your router.

Once the models are downloaded, none of this needs the internet. LM Studio's offline docs say chats, document chats and the local server all work without a connection, and nothing you type leaves your device. Only searching, downloading models and updates go online.

If it runs slowly or will not load

Most problems come down to memory. Work through these in order:

• "Failed to load" or a crash on load: lower GPU Offload, lower Context Length, or drop to a smaller model or a Q3 file.
• Painfully slow replies: the model is probably spilling out of VRAM into system RAM. Use a smaller model so it fits on the GPU.
• Whole PC freezes: you're out of RAM. Close the browser tabs and games, then load a smaller model.
• R1 model repeats itself or loops: set temperature to 0.6 and remove any system prompt on the older distills.
• GPU not being used at all: update your graphics driver, then check that a GPU runtime is selected. lms runtime ls lists what's installed.

If you'd rather try a lighter tool, I've covered the best Ollama models for 8GB RAM PCs. Ollama is more command-line, LM Studio is more point-and-click. Both use the same kind of model files under the hood.

FAQ

Is LM Studio free to use?

Yes. The app is free for personal use and for internal use at work under its terms. Some optional extra features can be paid, but downloading and running models locally costs nothing.

Can I run the full DeepSeek model on my PC?

Not realistically. DeepSeek V4 Flash needs around 156 GB of memory in LM Studio, and V4.1 Flash is even bigger. What runs on a normal Windows PC are the distilled R1 models, from 1.5B to 32B parameters.

Do I need an Nvidia graphics card?

No. LM Studio runs on the CPU if there's no suitable GPU, and it supports Windows on ARM laptops too. A GPU with 4 GB or more of dedicated VRAM is recommended because it makes replies much faster.

Is it safe to use DeepSeek locally?

Running it in LM Studio, your prompts stay on your PC. The model file doesn't send anything to DeepSeek. LM Studio only goes online to search, download and check for updates. The answers can still be wrong, so check anything important.

Which quantization should I download?

Q4_K_M for most people. It's a good balance of size and quality. Go up to Q6_K or Q8_0 if you have spare VRAM, or down to Q3_K_L only if the model won't fit otherwise.

Can I use Llama and DeepSeek models commercially?

Usually, yes. The DeepSeek R1 models are MIT-licensed. Llama 3.1 and 3.2 allow commercial use under Meta's community licence, but you need a separate licence from Meta above 700 million monthly users, and you must show "Built with Llama" if you distribute it.

If you get stuck on a model that won't load, tell me your RAM, GPU and the file you picked in the comments, and I'll point you to one that fits.