DeepSeek-V4-Pro with Ollama: Complete Setup Guide

What you are actually running

Before typing commands, lock this mental model:

• Ollama local API base: http://localhost:11434/api (official docs).
• Ollama model id for this flow: deepseek-v4-pro:cloud (official Ollama library entry).
• DeepSeek direct API base (OpenAI format): https://api.deepseek.com.
• DeepSeek model ids in docs: deepseek-v4-flash, deepseek-v4-pro.

So this guide is not about pulling V4-Pro full weights to your laptop. It is about getting reliable Pro access via either Ollama cloud routing or DeepSeek direct API.

Which path to choose first

Install on macOS, Windows, and Linux

Use this section if you are starting from a clean machine. The goal is simple: get Ollama running, verify local API responds, then test deepseek-v4-pro:cloud.

macOS (Apple Silicon or Intel)

1. Install Ollama from the official download flow on ollama.com and open the app once.
2. Open Terminal and verify CLI:

ollama --version

3. Check that local API is alive:

curl http://localhost:11434/api/version

4. Run model route check:

ollama run deepseek-v4-pro:cloud

If your shell says command not found, restart Terminal or re-login so PATH updates.

Windows (PowerShell)

1. Install Ollama from the official Windows installer on ollama.com.
2. Open PowerShell and verify CLI:

ollama --version

3. Confirm API endpoint:

curl http://localhost:11434/api/version

4. Run DeepSeek route:

ollama run deepseek-v4-pro:cloud

Windows environment variable (current session):

$env:DEEPSEEK_API_KEY="your_real_key_here"

Persist for future sessions (user-level):

setx DEEPSEEK_API_KEY "your_real_key_here"

Linux (Ubuntu/Debian/Fedora/Arch class)

1. Install with official script:

curl -fsSL https://ollama.com/install.sh | sh

2. Verify CLI and service response:

ollama --version
curl http://localhost:11434/api/version

3. Run DeepSeek route:

ollama run deepseek-v4-pro:cloud

If systemd service is not running after reboot, check:

systemctl status ollama

Cross-platform verification checklist

• CLI works: ollama --version
• API reachable: http://localhost:11434/api/version responds
• Model route works: ollama run deepseek-v4-pro:cloud returns answer
• No key in source code: use environment variables or secret manager only

Run DeepSeek-V4-Pro via Ollama

Install Ollama if needed, then run the cloud-tagged model listed in the official library page:

ollama run deepseek-v4-pro:cloud

If your environment is agent-first, Ollama also lists direct launch helpers:

ollama launch claude --model deepseek-v4-pro:cloud
ollama launch codex --model deepseek-v4-pro:cloud
ollama launch opencode --model deepseek-v4-pro:cloud
ollama launch openclaw --model deepseek-v4-pro:cloud

Model tags can evolve quickly. Recheck the live library page before hardcoding in scripts.

Call it from Ollama API

Once Ollama is running, use its default local API endpoint:

curl http://localhost:11434/api/chat \
  -d '{
    "model": "deepseek-v4-pro:cloud",
    "messages": [
      {"role": "system", "content": "You are a concise coding assistant."},
      {"role": "user", "content": "Review this function for edge cases."}
    ]
  }'

If you are migrating existing OpenAI-style clients, Ollama also supports compatibility endpoints at http://localhost:11434/v1/ in official docs.

Create DeepSeek API key

1. Open the DeepSeek key page: platform.deepseek.com/api_keys.
2. Create a new key for your project/workspace.
3. Store it in your secret manager or environment variable, never in source code.
4. Set local env var:

export DEEPSEEK_API_KEY="your_real_key_here"

For teams, prefer one key per service (or environment) instead of sharing one key across all apps. On Windows PowerShell, use $env:DEEPSEEK_API_KEY="..." for current session or setx for persistence.

First DeepSeek API call

DeepSeek docs show OpenAI-format chat completions with base URL https://api.deepseek.com.

curl https://api.deepseek.com/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer ${DEEPSEEK_API_KEY}" \
  -d '{
    "model": "deepseek-v4-pro",
    "messages": [
      {"role": "system", "content": "You are a helpful assistant."},
      {"role": "user", "content": "Give me a 5-step API migration checklist."}
    ],
    "thinking": {"type": "enabled"},
    "reasoning_effort": "high",
    "stream": false
  }'

Use OpenAI SDK with DeepSeek

Because DeepSeek API is OpenAI-format compatible, you can keep your OpenAI SDK call shape and just switch base URL + model id:

from openai import OpenAI

client = OpenAI(
    base_url="https://api.deepseek.com",
    api_key=os.environ["DEEPSEEK_API_KEY"],
)

resp = client.chat.completions.create(
    model="deepseek-v4-pro",
    messages=[{"role": "user", "content": "Write test cases for input validation."}],
)

print(resp.choices[0].message.content)

Thinking mode and effort controls

DeepSeek docs expose:

• thinking.type: enabled or disabled
• reasoning_effort: high or max

Practical rule:

• Use thinking disabled for high-throughput, low-complexity prompts.
• Use thinking enabled + high for general engineering analysis.
• Use thinking enabled + max for hard debugging, deep planning, or agentic chains where quality matters more than speed/cost.

Pricing and cost control checklist

DeepSeek pricing is billed per 1M tokens with input split into cache-hit and cache-miss lanes. The docs also note a limited-time 75% discount window for Pro (with explicit UTC end date) on the pricing page snapshot.

• Always read live table before committing budget.
• Separate prompt templates (stable prefix) from dynamic user payload to improve cache-hit behavior.
• Cap output tokens per endpoint by use case.
• Log reasoning usage because high/max effort changes spend profile.

If you need the full pricing breakdown already translated into workflow language, see DeepSeek V4 Pro API pricing guide.

Production security checklist

• Do not ship API keys in frontend code or mobile binaries.
• Keep provider keys server-side only.
• Use separate keys for dev/staging/prod.
• Rotate keys on schedule and immediately on leak suspicion.
• Mask prompts/responses in logs if they may contain PII or credentials.

Troubleshooting

model not found on Ollama

Check exact tag: deepseek-v4-pro:cloud. Do not assume deepseek-v4-pro (without suffix) is available locally.

401 / invalid auth on DeepSeek API

Verify DEEPSEEK_API_KEY, no extra spaces, and that your request uses Authorization: Bearer ....

Unexpected high bills

Inspect prompt length, output cap, and whether you are overusing reasoning_effort=max on low-value requests.

Related reading

• DeepSeek V4 preview: model split and migration timeline
• Run DeepSeek V4 locally on Apple Silicon (mlx-lm)
• Ollama complete guide (2026)
• Ollama MLX preview on Apple Silicon

Sources

• DeepSeek API Docs — Your First API Call
• DeepSeek API Docs — Models & Pricing
• DeepSeek API Docs — Create Chat Completion
• DeepSeek Platform — API Keys
• Ollama Docs — API Introduction
• Ollama Docs — OpenAI compatibility
• Ollama Library — deepseek-v4-pro

Frequently asked questions