What you get and what you need first
When you finish this guide, you will have DeepSeek-R1 8B in a chat window and a local server at localhost:11434. The program’s documentation says this server listens only to the computer itself by default.
I plan about one hour for the whole job; that is my estimate, not a measurement. The download itself is shorter: a 5.2 GB file takes about 7 minutes on a 100 Mbit/s line. The math is simple: 5.2 GB × 8 = 41.6 Gbit, divided by 0.1 Gbit/s, gives 416 seconds. The rest of the hour goes to checking the hardware, installing and asking the first questions.
Prepare three things in advance:
- Disk space. The Windows documentation says the program itself needs at least 4 GB, and AI files take tens to hundreds of gigabytes more.
- Normal user rights. The Windows installer needs no administrator rights, and on Linux you turn on the service with sudo commands.
- An NVIDIA or AMD graphics card, or a Mac with an Apple M chip. Without one, the work runs on the processor and answers come more slowly.
I advise you not to start with the full DeepSeek-R1, which has 671 billion parameters, but with a distilled version. A distilled version is a smaller model trained on the answers of the large one. According to the DeepSeek-R1 model page on Hugging Face of 22 January 2025, the developer released six of them. They have 1.5, 7, 8, 14, 32 and 70 billion parameters.
They are built on the open Qwen2.5 and Llama3. The R1-0528 release added one more, R1-0528-Qwen3-8B, built on Qwen3, and this is the version Ollama installs under the 8b tag. A distilled version runs on one GPUGPU. A GPU (graphics processing unit) is a graphics processor: a chip that performs thousands of identical calculations at the same time, so neural networks are trained and run on it. inside your own network, so running it is a sensible first step for a business. You need the full version only when the task demands its quality and you have a server.
This guide was checked against the Ollama and LM Studio documentation and the DeepSeek pages on Hugging Face on 20 November 2026. On that day the latest stable releases were Ollama 0.40.16 of 9 November and LM Studio 0.4.28. These programs change often: if a window or command looks different for you, trust their documentation. How to pick a version for your task is covered in my guide on which DeepSeek version to choose.
Which DeepSeek fits on your computer

DeepSeek fits on your computer when you have more video memory than the file size, plus room for context. In the Ollama library the 8B model is 5.2 GB, and an 8 GB GPU is enough for it.
| Tag | File | Memory with headroom | What I would use it for |
|---|---|---|---|
| 1.5b | 1.1 GB | any GPU or the processor | a test on a weak laptop |
| 7b | 4.7 GB | 8 GB | short answers and drafts |
| 8b, the default | 5.2 GB | 8 GB | the first working setup |
| 14b | 9.0 GB | 12 GB | contracts and reports with reasoning |
| 32b | 20 GB | 24 GB | hard tasks on one strong card |
| 70b | 43 GB | 48 GB | a workstation with a lot of memory |
| 671b, the full R1 | 404 GB | hundreds of gigabytes | a server |
| V4-Flash | about 160 GB | hundreds of gigabytes | a server, context up to 1 million tokens |
File sizes come from the same library and V4-Flash data from its Hugging Face page, both checked 20 November 2026. The memory and task columns are my rule from the file sizes, not a measurement. Tag sizes are for Q4_K_M, the default 4-bit quantizationLLM quantization. LLM quantization is a way to compress a neural network by storing its numbers with lower precision, so the model takes less memory and runs faster. that stores the model’s numbers compressed. The tags page on 20 November 2026 lists 8b in 8-bit q8_0 at 8.9 GB, near 14b, so weigh the two.
Here is the rule with its thresholds. You need 8 GB of video memory for 7b or 8b and 12 GB for 14b. For 32b you need 24 GB, and for 70b 48 GB. If memory is tight, take the row above, because a long document will use the rest.
A smaller size costs quality: DeepSeek tested its versions on AIME 2024, an American math olympiad (R1 model card, 22 January 2025). The 32B version solves 72.6% of the problems, 7B 55.5% and 1.5B 28.9%. The full R1 solves 79.8% in the same test, and these are the developer’s numbers, not an independent test.
Those results are for the January 2025 versions; the newer 8b scores 86.0 on the same test (Step 3). The DeepSeek-R1-0528 card on Hugging Face, which gives no date, puts the full R1-0528 at 91.4 on AIME 2024, above the small 8b.
For an exact count for the length of your documents, use the memory calculator. Why you need headroom beyond the file is explained in the analysis of how much memory a model needs.
Step 1. Check your system, graphics card and disk
Ollama runs on Windows 10 22H2 and newer, macOS Sonoma 14 and newer, and Linux, according to its documentation on 20 November 2026. Compare your system with the table:
| System | What you need | What speeds it up |
|---|---|---|
| Windows | Windows 10 22H2 or newer, Home or Pro; NVIDIA driver 551.61 or newer, and 570 or newer for GTX 10xx and other compute capability 5.0–6.2 cards | NVIDIA; AMD Radeon RX 7600 to 7900 XTX through ROCm; other cards through Vulkan |
| macOS | Sonoma (14) or newer | Apple M: processor and graphics through Metal; an Intel Mac uses the processor only |
| Linux | the install script from the official site; the CUDA or ROCm v7 driver is installed separately | NVIDIA through CUDA, AMD through ROCm, other cards through Vulkan |
Then check the hardware in this order:
- Find the name of your GPU and its memory size. On Windows, Task Manager shows both on the Performance tab; on Linux, the nvidia-smi command does.
- Look up the compute capability of your NVIDIA card in the documentation table. RTX 40xx is 8.9, RTX 30xx is 8.6, and GTX 1060, 1070 and 1080 are 6.1. You need 5.0 or higher.
- Check the NVIDIA driver. The Ollama hardware page says GTX 10xx and other cards with compute capability 5.0 to 6.2 need version 570 or newer. This applies on Windows too. Later cards, such as RTX 30xx and 40xx, need 550 or newer; the Windows install page asks for 551.61 or newer.
- Free up space on the system disk. On Windows the program goes into your home directory, and the 14b file adds another 9 GB.
Result of this step: you know which row of the memory table is open to you. A Mac with an Apple M chip shares memory between the processor and graphics, so the count works differently. The details are in the comparison of a Mac and a graphics card for running AI locally.
Step 2. Install Ollama
Ollama installs from an installer on Windows and macOS and with one command on Linux. Success looks the same everywhere: the ollama -v command prints the version. I choose this program because one install gives you both a chat and a local server, and they work the same on all three systems. Other options are covered in the comparison of tools for running AI locally.
On Windows:
- Download the OllamaSetup.exe installer from the official site and run it. You will not need administrator rights; the program goes into your home directory.
- Wait for the install to finish. The program starts in the background, and its command becomes available in cmd and PowerShell.
- Open PowerShell and check the version with the command from the start of this step. You should see the version number.
On macOS:
- Open the ollama.dmg image and drag the app into the Applications folder.
- Start the app. If its command is not yet on your system, the app asks for permission to create a link in /usr/local/bin. Allow it.
- Check the version in the Terminal app. You should see the version number.
On Linux:
- Run the install command from the official site in a terminal: curl -fsSL https://ollama.com/install.sh | sh.
- Make sure the service is running in systemd. The journalctl -e -u ollama command opens its log.
- Check the version with the same command. It prints the version number here too.
Result of this step: the version check passes, and on 20 November 2026 the command prints 0.40.16 or newer. Other tools your team uses will connect to localhost:11434 later.
Step 3. Download the model and ask the first question
The ollama run deepseek-r1 command downloads DeepSeek-R1-0528-Qwen3-8B, a 5.2 GB file, and opens a chat. That is how the Ollama library describes it on 20 November 2026.
- Type this command. For a different size, add a tag to the name, for example deepseek-r1:14b. After the download, a prompt for your question appears.
- Ask a work question from your own field that you already know the answer to, for example about the payment term in a standard contract. You see the reasoning first, then the answer.
- Compare the answer with what you know. Riddles and questions about the model itself say little about its value for work.
Why is 8B the default and not the older 7B? According to the DeepSeek-R1-0528-Qwen3-8B page on Hugging Face, it scores 86.0 on AIME 2024, against 76.0 for the plain Qwen3-8B. This is the developer’s own test, and the page gives no publication date.
The full R1-0528, whose reasoning this version learned from, improved on AIME 2025 from 70% to 87.5%, according to the same source. It also thinks longer: 23,000 tokenstoken. A token is a piece of text that a language model works with: a word, part of a word, or a symbol. per question on average instead of 12,000. That is why even the small version sometimes reasons for a while before it answers.
For the R1 series the developer recommends a temperature from 0.5 to 0.7, ideally 0.6 (R1 model card, 22 January 2025). Outside this range, the same card says, answers can repeat endlessly or lose coherence. In the chat, the /set parameter temperature 0.6 command sets it. The same source advises putting all instructions in the request itself, but the R1-0528-Qwen3-8B page says a system promptprompt. A prompt in AI is the text with which you give a task to an AI model: a question, an instruction, examples and the data the model uses to write its answer. is supported since version 0528.
If you downloaded the model earlier, the ollama pull deepseek-r1 command brings the latest version. Result of this step: DeepSeek answers your question, and you can see its reasoning before the answer.
Step 4. Make sure the model runs on the graphics card

The ollama ps command shows where the model is loaded, and the result you want in its PROCESSOR column is 100% GPU. The values below come from the program’s documentation:
| PROCESSOR column | What it means | What to do |
|---|---|---|
| 100% GPU | fully in video memory | nothing, this is normal |
| 100% CPU | in system memory, slow answers | install a newer driver or take a smaller model |
| 48%/52% CPU/GPU | the model did not fit in video memory | take a smaller model or shorten the context |
- Ask a question and right away type ollama ps in a second terminal window. Result: a line with the name deepseek-r1 and the PROCESSOR column.
- Compare the value with the table and do what its third column says.
- On Linux with an NVIDIA card showing 100% CPU, type nvidia-smi. The command should print details of your GPU, and an empty answer means the driver is not installed.
According to the program’s FAQ, the model stays in memory for 5 minutes after a request, so check it right away. The ollama stop deepseek-r1 command unloads it sooner and frees video memory for other programs.
Result of this step: the work runs on the GPU, and you can see how fast the answer is printed. To decide whether that performance is enough, read the guide to local speed and tokens per secondtokens per second. Tokens per second (TPS) is the speed at which a neural network writes its answer: how many pieces of text it produces in one second..
Step 5. Set the context, memory and file folder
According to the program’s context lengthcontext window. The context window is the amount of text, counted in tokens, that a language model sees at one time: instructions, documents, the question and its own answer. page on 20 November 2026, the default context window depends on video memory. Below 24 GB it is 4,096 tokens, from 24 to 48 GB 32k, and from 48 GB 256k. On an 8 or 12 GB card that is too small for a long contract or report. The context window is the amount of text the model holds at once: the document, the question and the answer.
- Enlarge the window for one session: in the chat, type /set parameter num_ctx 16384. Result: a longer document fits in the window.
- To keep the window for every start, set the OLLAMA_CONTEXT_LENGTH environment variable and restart the program. Each step up uses more video memory, so check the load again as in Step 4.
- Compress the context cache if the document does not fit. According to the program’s FAQ, setting OLLAMA_KV_CACHE_TYPE=q8_0 roughly halves its memory with almost no lossloss function. A loss function in machine learning is a formula that turns a model’s error into one number: the smaller it is, the closer the model’s answers are to the right ones. of precision. The q4_0 option cuts it to a quarter, with a small to medium loss.
- Compression works only with Flash Attention on. The program, the same page says, turns it on by itself when the hardware supports it. The OLLAMA_FLASH_ATTENTION=1 variable forces it on.
- Move the files to another disk if the system disk is small: the OLLAMA_MODELS variable sets the location. By default they sit in .ollama/models inside your home directory, and on Linux in the service’s system directory.
One server can serve several people, but parallel requests enlarge the context and the memory it uses. According to the same documentation, extra requests wait in a queue of 512, and beyond that the server returns a 503 error.
Result of this step: a document of the length you need is read in full. You also know how much video memory a second user at the same time will take. You can count the memory for your window in advance with the same calculator.
No command line: the same path in LM Studio
LM Studio gets you the same result without a terminal: search, download and chat in one window. Its requirements are stricter than those of Ollama, according to its requirements page on 20 November 2026:
- Mac: Apple Silicon only (M1, M2, M3, M4) and macOS 14.0 or newer. 16 GB of memory is recommended, 8 GB runs only small models with a short context, and Intel Macs are not supported.
- Windows: an x64 processor with AVX2 instructions or ARM (Snapdragon X Elite). At least 16 GB of system memory and 4 GB of video memory are recommended.
- Linux: an AppImage file for x64 and ARM64 and Ubuntu 20.04 or newer. Versions newer than 22 are not well tested by the developer.
Then follow these steps:
- Download LM Studio from lmstudio.ai and install it. Result: the program window opens.
- In the search tab, find DeepSeek-R1-0528-Qwen3-8B, the model the Step 3 command installs. Download its 4-bit Q4_K_M variant, as in the memory table. Result: the model appears in your downloaded list.
- Select this model and open the chat. Result: it answers the same work question as in Step 3.
- If other programs need DeepSeek, turn on the local server, which answers in the OpenAI format. Result: programs on this computer connect to it.
Once the model is downloaded, the program works with no internet. LM Studio says on its offline page that nothing you type into the chat leaves your device. You end up where Step 3 left you, only without the command line.
How to keep company data inside
By default Ollama listens only on 127.0.0.1, port 11434, which means the computer itself (program documentation, 20 November 2026). A colleague on the office network cannot connect until you open access using the OLLAMA_HOST variable.
I advise opening network access only through a VPN or a proxy server with a password. Close port 11434 to everyone else with a firewall. Otherwise anyone on the same network can reach the server.
The program’s FAQ says (checked 20 November 2026) its developer does not see your requests and data when you run locally. That is the developer’s own claim, so it is safer to turn off the cloud features too. Use the OLLAMA_NO_CLOUD=1 variable or the matching setting in server.json. After a restart, the log shows the line Ollama cloud disabled: true.
For the full R1, R1-0528 and V4-Flash, the MIT license allows commercial use. For R1 its card of 22 January 2025 says so, and for V4-Flash its own card. For R1-0528 the license is named on the card of its small version, R1-0528-Qwen3-8B, which puts the DeepSeek-R1 series under MIT and allows commercial use. The distilled versions also carry the license of their base modelfoundation model. A foundation model is a large neural network trained on a huge amount of general data, which teams then adapt to many specific tasks., covered in the questions below.
The main question here is not technical. I made this point in my talk on server crashes and new viruses. I told business owners to ask their developers or IT team how we can protect ourselves from data leaks into AI. A local modellocal LLM. A local LLM is a neural network that runs on your own computer or server, not in a vendor’s cloud, so your data goes nowhere. gives one of the answers.
I apply the same test to the contractor who installs DeepSeek for you. I put it this way in my talk on building software with AI. When you choose an outsourcing company to work on your product, your task is its security.
So start by putting your data in order. In the same talk on server crashes, I said you must start processing and structuring your data. Then more advanced models can work with it more effectively.
Running locally sends no requests to the model developer or its cloud, but by itself it does not make processing personal data lawful. Under Russia’s personal data law, Federal Law No. 152-FZ of 27 July 2006, a company still needs a lawful basis (Article 6). It also needs protection measures (Articles 18.1 and 19). Article 18, Part 5 (as amended by Federal Law No. 23-FZ of 28 February 2025) governs the databases used at collection.
This part covers collecting personal data of Russian citizens, including over the internet. At that stage, the data may not be recorded, systematized, accumulated, stored, updated or retrieved in databases outside Russia. The law allows exceptions only for the cases in Article 6, Part 1, Clauses 2, 3, 4 and 8.
Under Article 3, Clause 11 of the law, a cross-border transfer is a transfer of personal data to the territory of a foreign state. The recipient can be that state’s authority, a foreign individual or a foreign legal entity. So sending requests with this data to a cloud service outside Russia is exactly such a transfer.
If the computer or server that runs the model is inside Russia, there is no cross-border transfer. Sending data to a GPU server rented abroad is still a cross-border transfer: the data goes to another country and to a foreign company. For such a transfer, the operator must notify Roskomnadzor, Russia’s data protection regulator, of its intention before it starts (Article 12, Part 3). This notice is separate from the one under Article 22.
Article 12, Part 5 adds a step that comes before that notice. The operator must get the foreign recipient’s name and contacts, its protection measures and its terms for ending processing. If the country is not on Russia’s list of adequate-protection countries, Part 5 also requires details of that country’s data protection law.
Article 18, Part 5 forbids collecting the data straight into a database on a server abroad. A transfer abroad of data first collected into a database in Russia follows the rules of Article 12. Whether a company’s own server abroad, with no foreign recipient, counts as a cross-border transfer is disputed: Article 3, Clause 11 names foreign recipients. The rest is in the analysis of AI and personal data under the law.
Common mistakes and how to check the result
The seven mistakes in the table follow directly from the Ollama and LM Studio requirements on 20 November 2026, and each has a quick fix. Check memory first, because it fails before anything else.
| Mistake | Cause | Fix |
|---|---|---|
| The Step 4 check shows 100% CPU | the model is larger than video memory | take a smaller model from the memory table |
| The program does not see the GPU | NVIDIA driver below the Step 1 minimum: 550, 551.61 on Windows, and 570 for compute capability 5.0–6.2 cards | update the driver |
| The program does not work on Windows | the system is older than Windows 10 22H2 | update Windows |
| The file download breaks off | the home directory ran out of space | move the files using the OLLAMA_MODELS variable |
| An Intel Mac answers slowly | on x86 Ollama uses the processor only, and LM Studio does not support such a Mac | take 1.5b or 7b, or a Mac with an Apple M chip |
| The model repeats itself | the temperature is outside the recommended 0.5–0.7 range | set it to 0.6 |
| A long document is cut off | below 24 GB of video memory the default window is 4,096 tokens | enlarge the window as in Step 5 |
Three signs confirm the result:
- The command from Step 2 prints the version number.
- ollama ps shows 100% GPU while the model answers.
- The model answers a work question and correctly states a fact you know.
If the check fails, look for the cause in the logs. On Windows this is server.log in the program folder inside %LOCALAPPDATA%. On a Mac it is the logs folder next to the files, and on Linux the journalctl command from Step 2.
My threshold is this: if all three signs hold on ten work questions, you can give DeepSeek to your staff. If even one fails, go back to the step where it broke.
When the full DeepSeek or V4 is beyond your hardware
The full DeepSeek-R1 is 404 GB in the Ollama library (checked 20 November 2026), so no single card can hold it. The developer itself points to the DeepSeek-V3 repository for running it. The R1 model card of 22 January 2025 also says the Transformerstransformer. A transformer model is a neural network architecture that reads all the text at once. It uses a mechanism called attention to decide which words matter for understanding each word. library does not support it directly.
Even the 32B distilled version runs on two GPUs through vLLM, with a 32,768-token window, in the example from the same card. By my rule from the memory table, its 20 GB file fits on one 24 GB card only with a short window.
DeepSeek-V4-Flash has fewer parameters than the full R1: 284 billion, of which 13 billion are active for each token. That is what the V4-Flash page on Hugging Face says (checked 20 November 2026), and it is a preview version of the V4 series.
Running V4-Flash locally is harder. According to the same page, the weights use mixed precision: the experts in FP4 and most other parameters in FP8. The developer points to the inference folder with weight conversion and, for the Think Max mode, recommends a window of at least 384,000 tokens.
My decision is this: on one card with up to 24 GB, stay with the 8B to 32B distilled versions. A server-class version makes sense when the task needs the full R1’s quality and you have a server with hundreds of gigabytes. Otherwise, in my view, the cloud or the developer’s API costs less; see the guide to the DeepSeek API.



