How to install DeepSeek locally means installing a local inference runtime such as Ollama or LM Studio, downloading compatible DeepSeek-R1 weights, and running inference on your computer; you are not installing the hosted DeepSeek website or API. A quantized 7B/8B distilled model suits many 16GB systems, while full 671B R1 belongs on specialist multi-GPU hardware. For most beginners, Ollama is the fastest terminal-based route and LM Studio is the easiest graphical route. Developers who need a local service should use vLLM or SGLang with a distilled checkpoint. The instructions below separate those practical paths from full-model deployment so that a large download is not mistaken for a viable hardware requirement. Ollama is the simplest command-line route and supports macOS, Windows, and Linux. LM Studio is the simplest graphical route and supports local chat, model downloads, and local APIs on macOS, Windows, and Linux. A quantized 7B or 8B DeepSeek-R1 distilled model is the practical starting range for many computers with 16GB of system memory. Ollama’s listed DeepSeek-R1 tags range from approximately 1.1GB for 1.5B to approximately 404GB for 671B, but download size is not the same as required RAM or VRAM. The full DeepSeek-R1 is a 671B-parameter mixture-of-experts model intended for specialist multi-GPU or server deployment, not an ordinary laptop installation. What does installing DeepSeek locally actually mean? Installing DeepSeek locally means setting up an inference runtime and then downloading model weights to your computer. The runtime loads the weights, uses your CPU and/or GPU to generate responses, and optionally exposes a local API. The process does not install the hosted DeepSeek website, and it does not automatically provide access to DeepSeek’s hosted API. DeepSeek offers several model forms. The smaller DeepSeek-R1-Distill checkpoints are distilled models based on Qwen or Llama families and are suitable for common local runtimes. The full DeepSeek-R1 checkpoint is a very different deployment target. DeepSeek’s official R1 repository documents separate approaches for distilled models and large-scale inference. Which local DeepSeek runtime should you choose? Choose Ollama for the shortest terminal workflow, LM Studio for a graphical desktop workflow, and vLLM or SGLang when you need a developer-facing local API or server. For a first installation, use deepseek-r1:7b with Ollama or download a compatible 7B/8B distilled model in LM Studio. Move to a 14B or 32B model only after confirming that your system has enough memory for the model, context window, operating system, and runtime overhead. What hardware and model size do you need? The correct model size depends on quantization, context length, offloading, and whether the runtime uses system RAM, dedicated VRAM, or both. A model’s download size is only a storage requirement; the running process also needs memory for the operating system, runtime, KV cache, context, and temporary buffers. Ollama’s DeepSeek-R1 model catalog lists the following approximate sizes for its tagged downloads: Recommended Free Tools PC Slower Than It Used to Be? Crashes, No Sound, or Screen Glitches? Windows Errors? Fix Them Before They Spread DeepSeek AI’s official model card identifies full DeepSeek-R1 as having 671B total parameters, 37B activated parameters, and a 128K context length. The activated-parameter figure does not make the full model a 37B download: the model still contains 671B total parameters and requires infrastructure capable of holding and serving the complete checkpoint. LM Studio’s product guidance published on January 29, 2025, associates 16GB of RAM with smaller 7B/8B distilled models and describes the full 671B model as requiring roughly 192GB or more of RAM. Those figures are approximate practical guidance, not guaranteed minimums for every quantization, context length, operating system, or runtime. Platform requirements to check first LM Studio lists the platform requirements on its current system-requirements page . Ollama’s macOS documentation covers its platform-specific requirements and storage considerations. Crashes, No Sound, or Screen Glitches? Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely. How do you install DeepSeek locally with Ollama? Ollama provides the quickest command-line installation because a model can be downloaded and started with one command after the runtime is installed. Install Ollama. Use the official Ollama quickstart to choose macOS, Windows, or Linux. On Linux or macOS, the documented installation script is: curl -fsSL https://ollama.com/install.sh | sh Windows users should use the Windows installer described in the official quickstart rather than running the Unix shell command. Windows users should use the Windows installer described in the official quickstart rather than running the Unix shell command. Check available storage and memory. The deepseek-r1:7b tag is listed at approximately 4.7GB, but Ollama needs additional memory while the model is running. Close memory-heavy applications before loading a model if the computer is close to its limit. Download and start a practical model. Run: ollama run deepseek-r1:7b The first run downloads the model if the model is not already present, starts the local runtime, and opens an interactive terminal chat. The command uses a 7B distilled model, not the full 671B DeepSeek-R1. The first run downloads the model if the model is not already present, starts the local runtime, and opens an interactive terminal chat. The command uses a 7B distilled model, not the full 671B DeepSeek-R1. Try a different size only when the computer can support it. Ollama’s catalog lists 1.5B, 7B, 8B, 14B, 32B, 70B, and 671B DeepSeek-R1 tags. A smaller tag is the correct recovery step when a larger tag fails to load; downloading a larger file does not create additional RAM or VRAM. Use the local API when a script needs DeepSeek. Ollama exposes the local chat endpoint at http://localhost:11434/api/chat . A basic request has this shape: curl http://localhost:11434/api/chat -d '{"model":"deepseek-r1:7b","messages":[{"role":"user","content":"Explain local inference in one paragraph."}]}' The Ollama quickstart documents the endpoint and request structure. The Ollama quickstart documents the endpoint and request structure. How do you control Ollama’s context length? Ollama’s FAQ states that the default context length is 4096 tokens and that the OLLAMA_CONTEXT_LENGTH setting can change the context length. A larger context consumes more memory, so reduce the context length when a model runs out of memory or becomes unusably slow. On macOS or Linux, a session-specific example is: Use the syntax appropriate for the user’s shell and operating system. Increasing the context length is not a free quality setting: the runtime must allocate more memory for a larger context window. How do you install DeepSeek locally with LM Studio? LM Studio provides the simplest graphical workflow: install the application, find a compatible DeepSeek model, download the model, load the model into memory, and chat locally. Free tools Windows power users keep installed One-click scans. No signup required. Install LM Studio. Confirm the computer meets the current LM Studio system requirements . LM Studio documents support for macOS, Windows, and Linux, including x64 and ARM64 Linux systems. Open model discovery. Search for DeepSeek in LM Studio’s model catalog. Select a model whose format and size fit the computer. LM Studio’s DeepSeek R1 guidance discusses smaller distilled models for computers with around 16GB of RAM and distinguishes those models from the much larger full checkpoint. Download a compatible model. Choose a GGUF model for the relevant desktop backend or an MLX model where appropriate for Apple Silicon. Do not choose the full 671B model merely because the catalog displays it; check storage and memory before starting the download.

Load the model. Open the Chat tab and load the downloaded model into memory. LM Studio explains that loading allocates memory for model weights and other parameters. The model must fit with enough headroom for context and the operating system. Start a local conversation. After loading completes, enter a prompt in the Chat tab. The response is generated by the local runtime on the computer rather than by the hosted DeepSeek website. If a GGUF file already exists on the computer, LM Studio’s lms import documentation explains how to import the file into the local model directory. The import workflow supports moving, copying, hard-linking, or symbolically linking the existing file, depending on the chosen option and filesystem. Can LM Studio expose a local API? Yes. LM Studio documents a local REST API with an OpenAI-compatible interface, as well as a CLI and headless operation. The local API is useful when an application needs to send prompts to a model running on the same computer. Use LM Studio’s current API documentation and application labels because API controls can change between releases. LM Studio also documents offline operation after model files are available. The application can run inference locally without sending prompts to a remote service, but initial installers, model downloads, update checks, extensions, API integrations, and optional cloud features may still use the network. How do you serve a distilled DeepSeek model with vLLM? vLLM is the developer-oriented path for serving a DeepSeek distilled checkpoint through an OpenAI-compatible local server. vLLM changes quickly, so check the current vLLM CLI guide and installation matrix for the installed Python version, CUDA or ROCm stack, GPU architecture, and operating system. What’s actually slowing this PC down? Pick the symptom - the matching free tool is one click away. Start with the official 7B example DeepSeek’s official 7B model repository documents this basic workflow: The first command installs vLLM into the active Python environment. The second command downloads or locates the DeepSeek-R1-Distill-Qwen-7B checkpoint and starts a serving process. The official model repository documents an OpenAI-compatible chat-completions endpoint for the served model. Use a virtual environment in a real development setup, and follow the current vLLM installation instructions rather than assuming that the command above works for every GPU driver or Python environment. Serve the official 32B example across two devices DeepSeek’s official R1 repository provides this vLLM example for the 32B distilled model: The --tensor-parallel-size 2 setting is a two-device tensor-parallel configuration from the official example; the flag is not a universal instruction to use any two GPUs. The --max-model-len 32768 setting limits the serving context to 32,768 tokens, and the chosen context length affects memory requirements. The --enforce-eager flag is part of DeepSeek’s documented example and should not be changed casually without checking current runtime guidance. The current vLLM serve documentation describes options for a model name or local path, port, model format, download directory, and other serving parameters. Use those options when adapting the server to an existing model cache or an application that expects a particular port. Can you run the full DeepSeek-R1 model on a laptop? You generally should not treat the full DeepSeek-R1 as a laptop installation. The full checkpoint has 671B total parameters, an approximately 404GB Ollama catalog tag, and infrastructure requirements that go far beyond the practical 7B/8B local workflow. DeepSeek’s official V3/R1 repository documents distributed inference using multiple nodes and GPUs, along with engines including SGLang, LMDeploy, TensorRT-LLM, vLLM, and LightLLM. That deployment path is appropriate for a server or research cluster with carefully matched hardware and software, not for an ordinary 16GB laptop. A laptop may run a small distilled checkpoint from the DeepSeek-R1 family, but a smaller distilled model is not the same model as full DeepSeek-R1. Do not use the 1.5B, 7B, or 8B tag size as evidence that the 671B model will run on the same machine. What is the difference between DeepSeek-R1 and the distilled checkpoints? Full DeepSeek-R1 is the original 671B-parameter mixture-of-experts model, whereas DeepSeek-R1-Distill-Qwen and related tags are smaller checkpoints trained or distilled for more practical deployment. DeepSeek’s official R1 repository states that Hugging Face Transformers did not directly support the full R1 series at the time of that repository guidance, while the distilled models could be run like Qwen or Llama models. The distinction matters: do not paste a generic Transformers installation command into a full-model setup and assume that the command is universally valid. Check the current official repository before attempting a different framework. Is locally installed DeepSeek completely offline and private? Local inference can keep prompts and generated responses on the user’s computer after the runtime and model files are downloaded. Local execution is not the same as a guarantee that the entire installation process never connects to the internet. Crashes, No Sound, or Screen Glitches? Ollama needs network access to install the runtime and download model weights. LM Studio also needs network access to obtain the application and model files. Update checks, extensions, API integrations, and optional cloud features can create additional network activity. LM Studio’s offline-operation documentation explains the conditions under which the application can work offline, while Ollama documents local model execution in its official quickstart . For a stricter privacy setup, download installers and model files first, review application and firewall behavior, disable optional cloud integrations, and then disconnect the computer before testing inference. Those steps reduce network exposure, but local software behavior and operating-system telemetry still depend on the specific configuration. How do you troubleshoot a failed local DeepSeek installation? When LM Studio is installed through its CLI, the documented lms ps command can help show models currently loaded in the local environment. A loaded model that still fails to answer usually points to memory pressure, context settings, or a runtime/backend compatibility problem rather than a missing download. Which installation path is right for you? You want the fewest decisions: install Ollama and run ollama run deepseek-r1:7b . You want buttons and a chat window: install LM Studio, search for DeepSeek, download a compatible 7B or 8B model, load it in Chat, and start prompting. You need an application endpoint: use Ollama’s local API for a simple desktop workflow or vLLM for an OpenAI-compatible developer server. You have multiple GPUs and server experience: evaluate vLLM or SGLang with the official DeepSeek deployment instructions, and validate the exact model, context length, and parallelism settings. You mean the full 671B model: plan a specialist multi-GPU deployment rather than a normal laptop installation. The Bottom Line For most readers, the reliable answer to how to install DeepSeek locally is Ollama plus a quantized 7B or 8B distilled model, or LM Studio if a graphical interface is preferred. Confirm storage and memory before downloading, and treat the full 671B DeepSeek-R1 as a server-scale deployment. Crashes, No Sound, or Screen Glitches? High CPU Temps Kill Performance Leave a note Cancel reply Free Windows tools One Free Minute Could Fix That PC Before you go - each of these free tools takes about a minute and tackles what quietly slows a Windows PC down. Special offer. View Outbyte info, uninstall instructions , EULA , and Privacy Policy .