DeepSeek V4-Pro Won't Run on Your PC — Here's What Will
Install Ollama from ollama.com, then run ollama run deepseek-r1:7b. That pulls a 4.7 GB model file and gives you DeepSeek running entirely offline on your own machine.
But not the one in the headlines. DeepSeek's new V4-Pro is around 1.6 trillion parameters — there is no PC on earth that runs it, and there never will be.
That is the first thing nobody tells you this week. The second is stranger, and it changes what you think you are doing:
The "DeepSeek" most people run at home is not DeepSeek. The 7B and 14B models everyone downloads are distills — a different model, Qwen, trained to imitate DeepSeek's reasoning. It is free, it works offline, and it is genuinely useful. It is just not the thing the news is about, and almost nobody running it knows that.
The Graphics Card Bought for the Wrong Reason
A customer came into Jake's shop wanting a graphics card fitted. He had read that DeepSeek's newest model was competitive with the expensive American ones and free to run yourself, and he had decided he was done paying a monthly subscription for an assistant.
He had already bought the card. He wanted to know whether it was enough.
It was a perfectly good card. It was also never going to be enough, and neither would ten of them, because the model he had read about is measured in trillions of parameters and lives in a data center. What he could run was something quite different wearing a similar name.
The frustrating part is that he had not misunderstood anything he read. Every article was true. They were about a model release, and he was reading them as a hardware guide, and nothing in between said this one is not for your computer.
Ethan, when Jake described it: "The naming does the damage. There's a model in a data center and a much smaller thing you download, and they share a brand. Every headline is about the first one and every tutorial is about the second, and nobody writes the sentence that connects them. He spent real money on a reasonable inference from accurate information."
He kept the card, as it happens — it made the 14B model comfortable and it plays games. But the question he came in with deserved a straight answer, and this is it.
What DeepSeek Actually Released
Two things this month, and neither is a download.
V4-Pro reached general availability after months in preview. It is built for agent work — using tools, running code, completing multi-step jobs without someone watching. It handles a context window of up to a million tokens and can produce outputs as long as 384,000, and it runs in either a thinking or a non-thinking mode.
An experimental multimodal model followed on August 21, adding the ability to work from images and screenshots rather than text alone, positioned against the strongest models from the American labs.
The pricing moved at the same time, and it is worth knowing if you were planning to use the API: output pricing rose to around $3.96 per million tokens at peak hours, from a flat rate closer to $0.87, and DeepSeek introduced peak and off-peak billing with off-peak at roughly half the peak price.
None of that is something you install. These models run on DeepSeek's hardware and you reach them through the app, the website or the API. There are no weights to download, and even if there were, the file would be measured in terabytes.
Why the Big Model Will Never Run on Your PC
The arithmetic is simple enough to do in your head, and once you have done it the whole subject makes sense.
A model's parameters have to be held in memory while it runs. At the compressed formats used for home machines, roughly half a gigabyte of memory per billion parameters is a fair rule of thumb.
So a 7 billion parameter model needs somewhere near 4 to 5 GB. That fits on a modest graphics card. A 70 billion parameter model needs something in the region of 40 GB, which is data-center territory. A model in the low trillions needs hundreds of gigabytes of fast memory at minimum, spread across many cards that cost more than a car.
This is not a limitation anyone is going to optimize away. It is the size of the thing.
Which is why the number in the model name is the most useful piece of information in this entire subject. Learn to read 7b as "about 5 GB and it will fit," and you can predict what your machine can do without reading another article.
The Part Almost Nobody Knows: You Are Not Running DeepSeek
This is the bit worth the price of admission, and it is not a criticism of anything — it is simply not explained anywhere you would look.
When you run deepseek-r1:7b, you are not running a smaller version of DeepSeek R1. You are running a different model entirely.
What DeepSeek released alongside R1 was a set of distilled models. They took existing open models — Qwen from Alibaba, and Llama from Meta — and trained them on R1's reasoning output until they imitated its style of working through a problem. The result carries the DeepSeek name because DeepSeek made it, and it genuinely inherited something real. But the model underneath is Qwen or Llama.
| What you type | What it actually is | Download |
|---|---|---|
deepseek-r1:7b | Qwen, distilled from R1 | ~4.7 GB |
deepseek-r1:8b | Llama, distilled from R1 | ~5 GB |
deepseek-r1:14b | Qwen, distilled from R1 | ~9 GB |
deepseek-r1:32b | Qwen, distilled from R1 | ~20 GB |
| the actual R1 | 671 billion parameters | not for a PC |
| V4-Pro | frontier model, ~1.6 trillion | no weights released |
You do not have to take anyone's word for this. Once the model is installed, run:
ollama show deepseek-r1:7b
Read the architecture line. It says qwen2. Not deepseek2, which is what DeepSeek's own models report. The license header underneath says "Copyright (c) 2023 DeepSeek", because DeepSeek did the distillation — but the model being distilled into is Qwen, and the file tells you so in one line.
The same command also gives you the exact figures: 7.6 billion parameters, a context length of 131,072, quantized to Q4_K_M. Worth knowing, because those three numbers explain the memory it needs, how much you can paste into it, and why the download is 4.7 GB rather than 15.
Does this matter? Practically, less than it sounds. The distills are good, they are free, and they reason visibly in a way smaller models usually do not. Use them and be pleased.
It matters for one thing only: expectation. If you install a 4.7 GB file expecting the model from the benchmark charts and it disappoints you, you will conclude local AI is not ready. It is not that the local model failed. It is that it was never the same model, and nobody said so.
Choosing the Right Size for Your Machine
Find your hardware and take the matching row. Being ambitious here is the main way people end up with something unusably slow.
| Your machine | Run this | What to expect |
|---|---|---|
| No graphics card, 8 GB RAM | deepseek-r1:7b | Works, slowly. Ask and walk away. |
| No graphics card, 16 GB RAM | deepseek-r1:7b | Usable for drafting and summaries. |
| 6–8 GB graphics card | deepseek-r1:7b | Comfortable. The sweet spot for most people. |
| 12–16 GB graphics card | deepseek-r1:14b | Noticeably better answers, still quick. |
| 24 GB graphics card | deepseek-r1:32b | As good as home hardware gets. |
| A laptop with shared graphics | deepseek-r1:7b | Treat it as the no-graphics-card case. |
If you are not sure what you have, open Task Manager, go to Performance, and look at the GPU entry — dedicated graphics memory is listed there. Our guide to opening Task Manager covers the routes if the shortcut is not working for you.
Installing It
Two steps, and most of the time is waiting.
- Go to ollama.com, download the Windows installer, and run it. It installs as a background service — there is no window and nothing to configure.
- Open Command Prompt or PowerShell. No administrator rights needed, though our guide to running Command Prompt as administrator is there if something else needs it.
- Check it arrived:
ollama --version - Pull and run the model:
ollama run deepseek-r1:7b - Wait for the download. This is the part guides skip. It is 4.7 GB — the progress line shows your actual speed and an estimate, and both are honest. Do the arithmetic from the speed it reports rather than from anyone else's timing.
- When it finishes you get a
>>>prompt. Type a question and press Enter. Type/byeto leave.
Everything after that first download is fast, because the model is on your disk. Starting it again later takes seconds.
Useful once you have more than one: ollama list shows what is installed, and ollama rm deepseek-r1:7b removes one and gives the disk space straight back. Trying two sizes and keeping the one that suits your machine is a reasonable way to spend an evening.
When It Is Painfully Slow
Words appearing one at a time with long pauses is the most common complaint, and there are only really three causes.
It did not fit in video memory. By far the usual one. If the model is slightly too large for your card, it spills into system memory and the processor picks up the work, which is far slower. The symptom is distinctive: it runs, correctly, at perhaps a tenth of the speed you expected. Dropping one size down fixes it completely and the answers barely suffer.
You have no supported graphics acceleration. Then everything runs on the processor by design. That is not a fault and there is no setting to find. It is worth knowing before you spend an evening looking for one.
Something else is holding the graphics memory. A game, a browser with hardware acceleration and many tabs, video editing software. Close them and try again before concluding anything about your hardware.
Work through it in this order rather than reading forum posts about settings:
- Close everything else that uses the graphics card — games, video software, a browser with many tabs open.
- Start a reply, then open Task Manager and go to Performance.
- Watch the GPU and Memory graphs while it generates. GPU busy means it fitted; GPU near zero while Memory climbs means the work went to your processor.
- If it went to the processor, install the next size down —
ollama run deepseek-r1:7bif you were on 14b. - Compare the two on the same question. Keep whichever answers in a time you will actually tolerate.
Ethan's rule for this, which saves a lot of evenings: "Go one size down before you change any setting. A 7B model that fits will beat a 14B that doesn't, every single time, and it takes two minutes to test instead of two hours of reading forum posts about flags."
The Reason That Actually Justifies the Effort
If you want the best answers, the hosted models win and it is not close. So it is worth being clear about what running one locally is genuinely for.
Nothing you type leaves your computer. Once the file is downloaded you can unplug the network entirely and it keeps working, because there is no service to reach. No account, no logging, no terms of service that change next year, no question of what gets used for training.
That matters for a narrower set of people than the enthusiasm suggests, and it matters a great deal to them: anyone pasting client information, patient details, internal code, financial records or legal documents into an assistant. If your employer forbids the hosted tools, this is the version that does not break the rule.
The second reason is cost that does not move. DeepSeek raised V4-Pro output pricing to roughly $3.96 per million tokens at peak and split billing into peak and off-peak this month. That is their right and it is still inexpensive — but it is a reminder that hosted pricing is somebody else's decision. A model on your disk costs the same next year as today.
Jake keeps a 7B model on the shop machine for exactly one job: summarizing what a customer wrote on the intake form into something short before he starts. It never leaves the building, it costs nothing, and it does not need to be brilliant to save him ten minutes a day.
If you want more depth on the wider setup — other models, other tools, choosing between them — our guide to installing a local LLM on your PC covers the ground beyond DeepSeek specifically.
Your Questions, Answered Straight
Can I run DeepSeek V4-Pro on my own PC?
No. V4-Pro is a frontier model in the region of 1.6 trillion parameters, and there is no consumer hardware that will hold it. It runs in DeepSeek's data centers and you reach it through the app or the API. What you can run at home is a much smaller model that carries the DeepSeek name.
What is DeepSeek R1 7B that everyone runs locally?
It is a distilled model, and this surprises most people: it is not DeepSeek R1 made smaller. It is a different model, Qwen, trained to imitate R1's reasoning style. It is genuinely useful, it is free, and it runs offline, but it is not the model the headlines are about.
How much disk space does DeepSeek R1 7B need?
The download is about 4.7 GB. How long that takes is entirely your connection — a couple of minutes on fast broadband, over half an hour on a slow link. Leave several gigabytes of working space beyond the model file too, so plan for roughly 10 GB free before you start.
How much RAM or VRAM do I need?
The 7B model at standard quantization fits comfortably on 8 GB of video memory, and will also run on system RAM without a graphics card, much more slowly. The 14B model wants around 12 to 16 GB, and the 32B model is realistically a 24 GB card.
Do I need a graphics card to run a local model?
No, but you will feel its absence. Without a supported GPU the model runs on your processor and system memory, which works and is slow enough to change how you use it. It becomes something you ask a question and walk away from rather than a conversation.
Is running DeepSeek locally actually private?
Yes, and this is the strongest reason to bother. Once the model file is downloaded, nothing you type leaves your machine. You can disconnect from the internet entirely and it keeps working. That is a genuinely different privacy position from any hosted assistant.
Is the local model as good as the one in the app?
No, and it is not close. You are comparing a 7 billion parameter model to something several hundred times larger. The local one is capable for summarizing, drafting, explaining code and working offline. It is not the model that scored on the agent benchmarks.
What does the number after the model name mean?
It is the parameter count in billions, and it is the single most useful number in the whole subject. 7b means seven billion. It predicts the download size, the memory you need and roughly how capable the model is, all at once.
How do I install Ollama on Windows?
Download the Windows installer from ollama.com, run it, and it installs as a background service. Then open Command Prompt or PowerShell and type ollama run followed by a model name. There is no separate configuration step to complete first.
Why is my local model so slow?
Almost always because it did not fit in video memory and spilled onto the processor. The other two causes are a model too large for the machine, and other software already holding the graphics memory. Dropping to a smaller model fixes it more reliably than any setting.
Can I delete a model after trying it?
Yes. Run ollama list to see what is installed and ollama rm followed by the model name to remove one. The space comes straight back. Trying several and keeping the one that suits your hardware is the sensible approach.
Does running a local model cost anything?
Nothing beyond electricity and disk space. There is no account, no key and no usage billing. That is the other half of the appeal, and it is a real contrast with API pricing that can change without warning.
Should I use the API instead?
If you need the full model's capability, yes, because no local option approaches it. Be aware the pricing moves: DeepSeek raised V4-Pro output pricing and introduced peak and off-peak rates, with off-peak at half the peak price. Local costs stay flat forever.
Is LM Studio easier than Ollama?
It is friendlier if you would rather not use a terminal. It gives you a window, a model browser and a chat interface, and it downloads the same models. Ollama is simpler to script and lighter to run. Neither is wrong.
Will this work on Windows 10 as well as Windows 11?
Yes. Ollama supports both, and the commands are identical. What decides your experience is memory and graphics hardware, not which version of Windows you are running.
Can I run this on Linux too?
Yes, and the install is a single command from the Ollama site. Everything after that, the model names and the commands, is the same as on Windows. The models themselves are identical files.
Worth Reading Next
- The complete guide to installing a local LLM on your PC
The wider setup — other models, other tools, and how to choose between them. - How to disable Copilot in Windows
The other side of the same instinct — taking the AI you did not ask for out of your machine. - Open Task Manager in Windows
How to watch memory and GPU while a model runs, which is how you diagnose a slow one. - Run Command Prompt as administrator
Not needed for Ollama itself, but worth having for everything around it.
A note on this guide. Written August 22, 2026, the day after DeepSeek's experimental multimodal release and a few weeks after V4-Pro reached general availability. The download size here comes from actually pulling the model on a Windows machine rather than from the documentation. On the machine used for this page the 4.7 GB took 37 minutes at about 2 MB/s (i purpose fully tested on slow connection); the same file on a fast connection lands in a couple of minutes. Your speed is the only variable that matters, which is why this page gives you the size rather than a time. Model names and sizes move quickly in this field; if a command here does not match what you see, trust your screen and please write in so the page can be corrected. And if you arrived after reading about the new model and were hoping to run it yourself: you understood everything you read correctly. The articles were about one model and the tutorials are about another, and nobody writes the sentence in between. So keep testing folks! Don't worry, someone will make it run on our PC as well