Run DeepSeek Locally With Ollama

Ollama can run DeepSeek R1 and distilled R1 checkpoints locally on Windows, macOS and Linux. Install Ollama from its official site, choose an exact tag that fits your hardware, record the model digest and keep the local API bound to your own machine unless you deliberately add secure remote access.

Install Ollama

Use Ollama’s official download and documentation↗ — external site. On macOS, current requirements list macOS Sonoma 14 or newer, with Apple Silicon acceleration and CPU-only support for x86. Windows installation uses the official installer. Linux offers an install script and manual methods.

Review commands before running any shell installer. In managed environments, use approved packaging and verify the publisher. The cross-platform guide covers OS-specific storage and troubleshooting.

Choose a model tag

Ollama’s DeepSeek R1 library lists:

ollama run deepseek-r1:1.5b
ollama run deepseek-r1:7b
ollama run deepseek-r1:8b
ollama run deepseek-r1:14b
ollama run deepseek-r1:32b
ollama run deepseek-r1:70b
ollama run deepseek-r1:671b

The default deepseek-r1 tag may point to a current packaged variant rather than the original full R1. Use ollama show deepseek-r1:8b and record the digest, parameters, quantisation and template.

Start conservatively

Begin with the smallest model likely to meet the task. A 1.5B or 7B distill is easier to load but has less capacity than larger variants. The 671B model is not a normal laptop deployment.

Parameter size does not map directly to required memory. Quantisation, context length, runtime overhead and GPU offload affect the total. Ollama documentation notes that model storage can reach tens or hundreds of gigabytes. Check free disk before pulling.

Run a chat

After installation:

ollama run deepseek-r1:8b

Enter a bounded prompt, then type /bye to exit. Update an existing tag with:

ollama pull deepseek-r1:8b

An update can change behaviour. Pin or record the digest for reproducible evaluation and rerun your tests after pulling.

Call the local API

Ollama’s local API normally listens at http://localhost:11434. A simple request:

curl http://localhost:11434/api/chat \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-r1:8b",
    "messages": [{"role": "user", "content": "Explain two risks of generated code."}],
    "stream": false
  }'

No authentication is required for the local endpoint. That is acceptable only while it remains local and access to the machine is controlled.

Do not expose port 11434 casually

A process bound to all interfaces can let other devices use the model or retrieve data through connected tools. Keep the default local binding, use a firewall and never publish the port directly to the internet.

If remote access is necessary, place an authenticated TLS reverse proxy in front, restrict source networks, apply quotas and separate users. Do not rely on obscurity. Local model privacy depends on local endpoint security.

Files and sensitive data

Local inference can keep prompts off a hosted model service, but the client may write chat history or logs. Shell history can preserve commands. Backups can copy model prompts or attached files. Inspect the full application path.

Use a dedicated user account, restricted project folder and disk encryption where appropriate. Remove secrets from prompts. A local model connected to unrestricted filesystem tools can still expose private data.

Improve reliability

Give the model explicit output constraints and ask for checkable results. For code, run tests and security scans. For facts, use authoritative local documents or a controlled retrieval system and include source locations.

Small distills can loop or over-reason. Limit generated tokens, reduce context and split tasks. Do not increase every setting at once; measure one change against a fixed evaluation set.

Update and rollback

Record Ollama version and model digest before an update. Keep a small regression suite of real prompts and expected properties. If behaviour degrades, restore a known model artifact or tag when your deployment process supports it.

Follow Ollama’s release notes and DeepSeek model cards. A runtime update and a model update are separate changes.

Ollama versus the DeepSeek API

Ollama R1 tags run open checkpoints, often quantised or distilled. They are not the current V4 Flash or Pro hosted service. They may lack the one-million-token context, web search, account sync or provider tool behaviour.

Use the model overview to label results correctly. Compare exact versions, not “local DeepSeek” against “DeepSeek” as generic brands.

Local tags are not first-party API identifiers. Compare the hosted workflow in the DeepSeek API guide and the isolated Docker deployment path before choosing an operational model.

Conclusion

Ollama offers a practical route to local DeepSeek R1 use when the selected checkpoint fits your hardware. Install from official sources, record exact artifacts, protect the unauthenticated local endpoint and verify output with normal engineering or research controls.