To run DeepSeek R1 locally on Windows with Ollama, install the Windows application, choose a model your computer can handle, and launch it from PowerShell. Ollama manages the model download and local execution; this workflow does not require Python, Docker, WSL, or DeepSeek’s hosted API.
The walkthrough below uses DeepSeek-R1-0528-Qwen3-8B, a smaller distilled model—not the full 671B model. It also includes a lighter starting option, commands to verify the loaded model, and Windows-specific troubleshooting. The official Ollama quickstart distinguishes local use from cloud access: you do not need a cloud account or API key for this local setup.
Last checked: October 9, 2026.
Before downloading: check Windows, memory, and model size
Ollama’s Windows documentation lists Windows 10 22H2 or newer, Home or Pro, and at least 4 GB of disk space for the application’s binary installation. Model files require additional storage. Application compatibility is separate from Windows security support; use an appropriately maintained Windows installation.
This walkthrough targets Intel/AMD x64 Windows PCs. Ollama also publishes a separate Windows ARM64 package in its official releases, but do not assume that x64 GPU instructions apply unchanged to an ARM-based computer.
Check your processor, installed RAM, and system type under Settings → System → About. Microsoft documents the system-type check. For current memory usage, open Task Manager and select Performance → Memory; its performance views help distinguish installed capacity from memory already in use.
A dedicated GPU is not mandatory: Ollama can run through its CPU backend. For acceleration, check your exact GPU and drivers against the current hardware support documentation, including the Windows-specific AMD requirements. Vulkan provides another supported backend when available; neither an NVIDIA, AMD, nor Intel brand name alone guarantees that a particular configuration will work.
Which DeepSeek R1 model should you choose?
Distillation and quantization are different. A distilled model is a smaller model trained using reasoning examples from a larger model. Quantization reduces the numerical precision used to represent model weights. Neither means that a small download reproduces every capability of the full model.
DeepSeek’s official 8B model card identifies this release as a Qwen3-8B model post-trained using reasoning from DeepSeek-R1-0528. It should not be confused with the older Llama-based R1 8B distill.
| Ollama tag | Model identity | Listed download size | Suggested use |
|---|---|---|---|
deepseek-r1:1.5b | R1 Qwen-based 1.5B distill | 1.1 GB | The smallest starting experiment when memory is limited or you are trying CPU-only operation. |
deepseek-r1:8b | R1-0528 Qwen3-8B distill | 5.2 GB | The model used in this walkthrough, provided your computer has sufficient available memory. |
deepseek-r1:14b | R1 Qwen-based 14B distill | 9.0 GB | A larger model to evaluate after establishing a working setup and checking memory headroom. |
deepseek-r1:32b | R1 Qwen-based 32B distill | 20 GB | A higher-memory experiment, rather than a default first download. |
Model identities and download sizes come from the Ollama R1 library and its tag registry. The suggestions are starting recommendations, not official RAM requirements or performance guarantees. The full deepseek-r1:671b entry is listed at approximately 404 GB for its default quantization—a very different resource commitment.
A 5.2 GB download does not mean 5.2 GB of RAM or VRAM is sufficient. Running the model also needs working memory, including its context cache, while Windows and other applications retain their own memory requirements. Larger context settings increase memory demand, as explained in Ollama’s context-length documentation.
When your available memory is uncertain, start with 1.5B and a short prompt. Move to 8B only after checking whether the smaller setup leaves adequate headroom. Do not choose a larger model solely because its download fits on your drive.
Install Ollama and run DeepSeek R1 on Windows
1. Install the native Windows application
Open the official Windows download page and use the manual download option to obtain OllamaSetup.exe. Run the installer, then open Ollama from the Start menu. Choose local use rather than signing in for cloud models when completing setup.
Open Windows PowerShell from Start and enter:
ollama --versionYou should receive version information. A warning that Ollama cannot connect to a running instance means the command is installed but the local application or server needs attention; it is not evidence of a model-download problem.
The Windows application normally runs Ollama in the background. Do not start an additional ollama serve process during this standard installation path. Ollama documents this background behavior and its local server address in the Windows setup guide.
2. Download and launch the selected model
At the check date, deepseek-r1, deepseek-r1:latest, and deepseek-r1:8b point to the same registry entry. To make the release and quantization explicit, this walkthrough uses deepseek-r1:8b-0528-qwen3-q4_K_M.
Run this command in PowerShell:
ollama run deepseek-r1:8b-0528-qwen3-q4_K_MWhen the model is missing, ollama run downloads it before opening the conversation. A separate ollama pull step is therefore optional, not another mandatory installation stage. This behavior is implemented in Ollama’s command-line code.
Allow the download and loading stages to finish. The interactive prompt, typically displayed as >>>, indicates that the session is ready for your message.
Lighter alternative: to begin with the smaller model instead, use:
ollama run deepseek-r1:1.5bChoose one starting route; you do not need both downloads. The remaining model-specific commands use the explicit 8B tag. Substitute deepseek-r1:1.5b when that is the model you selected. Tags can change over time, so retain the model ID alongside the tag when recording results.
3. Send a first prompt with a checkable answer
Enter the following inside the model conversation—not at the ordinary PowerShell prompt:
A box contains 18 red cards and 12 blue cards. How many cards are there in total, and what fraction are blue? Give the total and the simplified fraction.The independently calculated answers are 30 cards and 2/5 blue. These are reference answers, not a claimed capture of this model’s output. A response demonstrates that the workflow produces text; one correct arithmetic answer is not a benchmark or proof of general reliability.
You may see thinking output before the final answer. Ollama separates reasoning output from answer output for supported models, and its thinking documentation explains that controls differ by model. Hiding a displayed reasoning trace is not the same as removing its computational cost.
4. Verify the model and its CPU/GPU placement
Open a second PowerShell window. Do not type these diagnostic commands into the model’s conversation:
ollama list
ollama show deepseek-r1:8b-0528-qwen3-q4_K_MCheck the model name and ID in the list, then inspect its metadata. The selected registry entry identifies the architecture as qwen3 and the quantization as Q4_K_M. Use this metadata rather than asking the chatbot to identify itself.
While the model is loaded, run:
ollama psAccording to Ollama’s GPU-loading guidance, 100% GPU means the model is loaded entirely into GPU memory; 100% CPU means system memory; a mixed CPU/GPU value means the model is split between them. These percentages describe model placement, not live GPU utilization.
An empty running-model list can simply mean the model has unloaded after inactivity. Send another message, then repeat the check. Record the allocated CONTEXT value too—it helps explain memory differences between otherwise similar setups.
Optional: use Ollama’s Windows chat interface
You do not need a third-party interface just to avoid terminal conversations. Ollama’s official Windows application includes a chat interface. Open it from Start, choose the local model you downloaded in the model selector, and enter your message.
Keep the CLI available for inspection and troubleshooting. A model available in a graphical interface is not automatically local: check the selected model rather than assuming every option uses your computer.
Keep the workflow local and verify offline operation
Downloaded model weights and cloud-hosted models are different execution paths. Ollama’s cloud documentation explains that cloud requests run on its servers and can be accessed through the app or terminal. For this guide, remain with the downloaded R1 tag and decline any suggestion to switch to cloud execution.
To disable Ollama’s cloud models and web-search features, set the documented OLLAMA_NO_CLOUD environment variable to 1. The local-only configuration instructions require the application to pick up the changed environment:
- Quit Ollama from its system-tray menu.
- Search Windows Settings for environment variables, open Edit environment variables for your account, and add a user variable named
OLLAMA_NO_CLOUDwith the value1. - Save the change and relaunch Ollama from Start. Open fresh PowerShell windows for subsequent commands.
Setting a variable only inside a new PowerShell window does not reconfigure an already-running tray application. Also check any previous customization: an OLLAMA_HOST pointing to another computer means your CLI is contacting that computer, not necessarily performing inference on yours.
For a practical offline check, first finish the download and obtain a response. Exit the conversation, unload the model using the commands below, disconnect the computer’s network connections, then run the same local model again and submit a new prompt. This checks the downloaded model’s offline workflow; it is not a complete security audit.
Installation, downloads, and updates still need connectivity. An offline model also cannot fetch current web information by itself. External interfaces and tools can change where data goes, so local execution should not be described as an unconditional privacy guarantee.
Keep the service on its default loopback address for this single-computer setup. There is no need to expose a port publicly or disable Windows Firewall to chat with a model on the same machine.
Fix common Windows errors and slow responses
Identify the failing stage before changing settings: command detection, connection, download, model loading, or generation. A problem in one stage does not automatically justify reinstalling everything.
| Symptom | Check first | Next action |
|---|---|---|
ollama is not recognized | Was PowerShell already open during installation? | Open a fresh PowerShell window. If necessary, confirm that Ollama’s installation directory was added to your user PATH. |
| Cannot connect to Ollama | Is the application running, and has the host address been customized? | Open Ollama from Start and verify the intended local connection. Do not start multiple servers blindly. |
| Port 11434 is already in use | Did you run ollama serve while the application was already running? | Stop launching the duplicate server and use the existing application. Do not terminate an unidentified process. |
| Model or manifest not found | Does the complete tag match the official registry? | Copy the verified tag again. A nonexistent model name is not fixed by adding RAM or changing drivers. |
| Download fails or storage fills up | Available space in the actual model directory and any network or proxy error | Resolve the reported issue, then retry the same download. On a managed network, ask its administrator about the documented proxy configuration. |
| Model fails to load or exhausts memory | Model size, other applications, and allocated context | Try 1.5B, close unnecessary applications, and reduce an oversized context setting before retrying. |
| Unexpected CPU use or slow generation | ollama ps, GPU compatibility, drivers, and server logs | Check the supported backend for your exact hardware. A smaller model may be more practical than partial CPU/GPU offloading. |
Use the official tags, hardware support page, and troubleshooting documentation for the corresponding checks. Avoid Linux-only GPU overrides, registry hacks, or security-setting changes copied from unrelated tutorials.
For the standard Windows installation, inspect the recent server log from PowerShell:
Get-Content "$env:LOCALAPPDATA\Ollama\server.log" -Tail 50Look for a specific loading, driver, memory, or connection error. If the file is missing, confirm that the application has started and consult the documented log locations. Remove private information before sharing logs publicly.
Reduce context when a larger setting is consuming memory
The model’s advertised context capacity is not the amount necessarily allocated in your session. Ollama’s dedicated context page documents hardware-dependent defaults, while its FAQ still describes a 4,096-token default. Check the actual CONTEXT column in ollama ps instead of assuming either number applies universally.
For a short-chat experiment using a larger context, enter this inside the running model conversation:
/set parameter num_ctx 4096Send another prompt and recheck the allocation. This is a session setting, not the model’s maximum context or a universal performance recommendation. It will not reduce memory if you were already using that context size. When the model cannot load at all, choose a smaller model or lower the application’s context slider before retrying.
Exit, unload, update, or remove the model
Exit the conversation without deleting anything
After a response finishes, enter the interactive exit command at the model prompt:
/byeThis returns you to PowerShell. It does not remove the downloaded model or quit the background application.
Unload the model from memory
At the PowerShell prompt, use the model-stop command:
ollama stop deepseek-r1:8b-0528-qwen3-q4_K_MThis unloads the model without deleting its files. Use ollama ps to check the result. To start another conversation later, repeat the original ollama run command; the downloaded files remain available.
Update Ollama separately from the model
For the application, use Restart to update when offered in Ollama’s tray menu, or install a current official Windows release. Ollama describes this process in its update instructions.
To check the registry for an updated copy of the selected model tag, run:
ollama pull deepseek-r1:8b-0528-qwen3-q4_K_MUpdating the application does not automatically mean you have refreshed every downloaded model. Likewise, pulling this tag does not switch you to a different DeepSeek family.
Remove a model you no longer need
Deletion command: the following removes the selected local model entry. Use stop, not rm, when you only want to release working memory.
ollama rm deepseek-r1:8b-0528-qwen3-q4_K_MConfirm the result with ollama list. Running a removed model again may require downloading its data. These commands are documented in the Ollama CLI reference.
Store future downloads on another drive
The default Windows model directory is under your user profile at %USERPROFILE%\.ollama\models. To use another location, follow the Windows model-location instructions: quit Ollama, create or select a writable folder, and set the user environment variable OLLAMA_MODELS to its full path.
For example, D:\OllamaModels is suitable only when that drive and folder exist on your computer. Save the setting and restart Ollama. Changing the variable does not itself move existing downloads: an empty destination may show no installed models until you download them there. Keep the original files until the new location is verified.