Run DeepSeek on Windows, macOS and Linux

The simplest cross-platform way to run a DeepSeek-family model locally is a supported runtime such as Ollama. Windows, macOS and Linux can all run smaller R1 distills, but GPU support, installation, storage paths and performance differ. No single hardware promise applies across models or quantisations.

Choose the model before the operating system steps

Ollama packages DeepSeek R1 distills from 1.5B through 70B plus full 671B. Start with a small tag that fits available memory and storage. An 8B distill is not the same model as full R1.

Quantisation reduces memory and download size but can affect quality. Context length also consumes memory. Record tag, digest, quantisation and configured context in every evaluation.

Windows

Ollama provides a Windows installer and normally installs without administrator privileges. Its documentation says the binary installation needs at least 4GB, with additional model files potentially using tens or hundreds of gigabytes. The model directory can be changed through OLLAMA_MODELS.

Use current GPU drivers. Windows GPU support varies by NVIDIA, AMD and integrated hardware. If a model falls back to CPU, confirm driver compatibility and check logs. Avoid random “DeepSeek for Windows” executables; no verified first-party DeepSeek desktop binary was found in the desktop app review.

macOS

Current Ollama requirements list macOS Sonoma 14 or newer. Apple M-series systems support CPU and GPU use; x86 Macs are CPU-only according to the documentation. Install the official application and allow it to place the CLI in the path.

Unified memory helps Apple Silicon share capacity between CPU and GPU, but the operating system and other applications also need memory. Close heavy workloads for large models and start with a smaller distill. Docker Desktop on macOS does not provide Ollama GPU passthrough, so native installation is usually the practical route.

Linux

Linux offers the broadest server and container options. Ollama provides an install script, manual packages and a system service path. NVIDIA acceleration needs compatible drivers; current AMD documentation references ROCm 7 for supported hardware.

Verify nvidia-smi or the relevant AMD tooling before blaming the model. Review service permissions and logs. A system service listening beyond localhost needs firewall and authentication design.

Basic commands

After installing Ollama:

ollama pull deepseek-r1:8b
ollama show deepseek-r1:8b
ollama run deepseek-r1:8b

Use /bye to leave. Test the API at http://localhost:11434 only from the local machine. The Ollama walkthrough provides a safe curl example.

Storage planning

Keep model storage on a fast local disk with enough headroom for updates. Downloads can leave temporary files. Backups of model weights may be unnecessary because they can be re-downloaded, but configuration and evaluation records may matter.

Changing OLLAMA_MODELS requires the Ollama service or app to have permission to the new location. Do not point multiple runtimes at the same mutable directory without documentation support.

Memory and GPU considerations

Model file size is a starting clue, not a guarantee that the model fits. Runtime buffers and context add memory. Some layers can be offloaded to a GPU while others remain in system RAM. Insufficient memory can cause slow swapping, process termination or load failure.

Begin with short context and one user. Increase settings incrementally while watching memory. Do not publish a tokens-per-second number without exact hardware, software, model and prompt details.

Platform differences

Windows offers straightforward desktop installation and broad consumer GPU use. macOS offers a simple native app and efficient Apple Silicon integration. Linux provides server automation, containerisation and the widest tuning surface, but requires more administration.

Model output should be broadly comparable when the exact artifact and settings match, though runtime kernels can produce small differences. Operational convenience often decides the platform more than model capability.

Common troubleshooting

If ollama is not found, restart the terminal and check the installation path. If a pull fails, verify disk space, network and proxy configuration. If generation is slow, confirm GPU detection, model size and context. If the service is unreachable, check whether Ollama is running and whether another process uses port 11434.

On Windows or Linux, outdated AMD drivers can cause GPU discovery problems. On macOS, unsupported operating-system versions or Intel CPU-only execution can limit performance. Consult current Ollama logs and hardware support rather than applying unrelated fixes.

Security and privacy

Local inference can keep prompts on the machine, but the local API has no authentication. Keep it bound to localhost. Do not open firewall ports or set broad origins unless a secure design requires it.

Restrict file permissions, logs and clients. An editor extension connected to local Ollama may still send telemetry or repository content elsewhere. Review every component.

When to use Docker

Docker is useful on Linux servers, CI or reproducible shared infrastructure. Bind its published port to 127.0.0.1, use a named model volume and grant only required GPU devices. Native installation is simpler for many individual desktop users.

See the Docker guide for secure commands.

A local runtime and the hosted service solve different problems. Compare the DeepSeek API guide when managed inference, current first-party models or simpler operations matter more than keeping inference on one machine.

Conclusion

Windows, macOS and Linux can all run suitable DeepSeek open models. Choose a checkpoint that fits, install a trusted runtime, plan storage and validate GPU detection. The exact model and security configuration matter more than the operating-system logo.