When people search for “running DeepSeek locally,” many imagine moving the high-scoring, full-size model onto a home computer and using it without an internet connection. Here is the short answer: almost every DeepSeek model a consumer PC can run locally is a distilled version—a much smaller model derived from the original. The full 671B-class model is impractical without a fleet of professional GPUs. This guide explains which versions you can run, how much VRAM they require, how to install them, and when you should skip the local setup and use the official website instead. The goal is to keep you from spending half a day on an installation only to discover that the result is worse than opening the site on your phone.
First, Decide Which Version You Actually Want to Run
DeepSeek has released models in several size classes, and the gap between them is enormous. One group consists of full, high-parameter-count models. They are the most capable, but they can occupy hundreds of gigabytes and require multi-GPU servers. They are outside the reach of most individuals. The other group consists of smaller distilled models, ranging from roughly one or two billion parameters to more than 30 billion. These are streamlined specifically so they can run on consumer hardware. A model installed on a personal laptop or desktop is almost always in the second group.
Do not pretend that the two groups are equivalent. A distilled model is sufficient for everyday Q&A, text cleanup, and some local writing, but it is not in the same class as the full model for complex reasoning, long documents, or professional coding. Before installing anything, ask why you want a local model. Is it to keep data private, work offline, or avoid API fees? If your only goal is to use the strongest version of DeepSeek, there is no need to deploy it yourself. The official site and API provide the full model, usually with better speed and accuracy. Clarifying the requirement first eliminates much of the unnecessary work.
Step 2: Choose a Model Size Based on VRAM
For local large-model inference, the hard limit is GPU memory, or VRAM—not CPU performance or system RAM. As a rough guide, a quantized 7B–8B model needs about 6–8 GB of VRAM and can run on a midrange gaming GPU. A 14B model generally needs 10–12 GB. To run a 32B model smoothly, you will usually want a card with 24 GB of VRAM. What happens when there is not enough? The model either fails to load or offloads part of its work to system memory and the CPU, slowing output to a word-by-word crawl.
A discrete GPU is not strictly required for experimentation. Apple's M-series chips use unified memory and perform surprisingly well with small and medium models; systems with more memory can even handle larger ones. CPU-only inference also works, but it is best reserved for trying the smallest 1.5B-class models. Do not expect it to support serious daily use. The practical rule is simple: check how much VRAM or unified memory your system has before choosing a download. Do not start with 32B just because it sounds better.
Step 3: Install It with Ollama or LM Studio
Beginners have two straightforward options. If you prefer a graphical interface, use LM Studio. Search for a distilled DeepSeek model, click Download, and chat with it in the application when the download finishes; the experience is similar to using a web app. If you prefer a minimal tool or want to connect the model to another program, use Ollama. After installation, one command downloads the model and starts a conversation. Ollama also exposes a local API, which is useful later for connecting a knowledge base or a small application of your own.
One important choice during installation is quantization. The same model may be available at several quantization levels. Lower-bit versions use less disk space and VRAM, but they also lose more quality. For most people, a middle option such as Q4 provides a reasonable balance between size and accuracy. Go lower only when memory is especially tight. Check the file size before downloading so you have a realistic idea of whether the model will fit.
Step 4: Put the Local Model to Useful Work
Getting the model to chat is only the beginning. The real value of local deployment comes from connecting it to your own materials. Ollama's local API can be paired with tools such as AnythingLLM or Dify. You can add PDFs, notes, and internal documents for question answering while keeping the data on your own machine. That is the central advantage over a web service: sensitive material does not have to be uploaded to a third party.
Keep your expectations realistic. Distilled models make mistakes and can state invented information with complete confidence, especially when a task involves numbers, dates, or professional judgment. Good uses include organizing messy text, extracting key points from meeting notes, drafting an email locally, and retrieval-based Q&A when your private documents contain an explicit answer. Think of the model as an offline assistant that keeps a secret but occasionally gets confused. It is valuable when assigned the right work.
Common Problems and Troubleshooting
The most common complaint is that the model becomes extremely slow after installation. In nine cases out of ten, there is not enough VRAM and part of the model has spilled into system memory. Use a smaller model or a lower quantization level instead of forcing an oversized model to run. Another common problem is a download that stalls or repeatedly disconnects. Model repositories can be slow or unreliable from within China, so configuring a suitable mirror or changing the network environment in advance can save time. A third mistake is expecting the local model to match the full version, then concluding that DeepSeek is weak. A distilled model is not a fair proxy for the full one. A fourth problem is an incorrect GPU driver: the application runs on the CPU because the driver or required compute environment is missing. Verify both before installation. When an error appears, search for the exact message first; most local-deployment errors have already been encountered and documented by someone else.
When to Skip Local Deployment and Use the Full Model
In practical terms, most people do not need a local deployment. If you want DeepSeek's strongest capabilities, the official web app and API use the full model. They respond quickly, consume none of your hardware resources, require no maintenance, and handle demanding work such as coding and complex analysis more reliably than a local distilled version. Local deployment is genuinely worthwhile in three situations: the data is highly sensitive and cannot be uploaded, the model must work without a network connection, or usage is frequent and heavy enough that local operation is cheaper than a metered API.
The most sensible arrangement is often a hybrid one: use the official site or API for everyday work and demanding tasks, and reserve a local distilled model for private files and offline emergencies. Decide which group your needs fall into before spending an afternoon configuring the environment. Using each tool where it is strongest is more practical than insisting that everything run on your own machine.