DeepSeek V4
Use DeepSeek V4 Flash and Pro directly from your terminal — one-shot questions, deep reasoning with thinking mode, and multi-turn chat. No special CLI required; uses the OpenAI-compatible API via a small Python script.
Setup
1. Get API key: https://platform.deepseek.com/api_keys
2. Set environment variable:
export DEEPSEEK_API_KEY=your_key_here
# Add to ~/.zshrc or ~/.bashrc to persist
Models
| Model | ID | Best for | Price (input/output) |
|---|---|---|---|
| V4 Flash ⚡ | deepseek-v4-flash | Q&A, writing, coding, summaries | $0.014 / $0.028 per 1M |
| V4 Pro 🚀 | deepseek-v4-pro | Hard reasoning, math, deep analysis | $0.174 / $0.348 per 1M |
Both support 1M token context. Cache hits are 10× cheaper.
Legacy aliases (deepseek-chat → flash, deepseek-reasoner → pro) deprecated 2026-07-24.
Commands
One-shot question (Flash — fast & cheap)
uv run {baseDir}/scripts/ask.py "Explain the difference between V4 Flash and V4 Pro"
One-shot with Pro model
uv run {baseDir}/scripts/ask.py "Write a merge sort in Rust" --model pro
Thinking mode (Pro with visible reasoning trace)
uv run {baseDir}/scripts/ask.py "Prove that there are infinitely many primes" --think
Multi-turn chat
uv run {baseDir}/scripts/chat.py --model flash
uv run {baseDir}/scripts/chat.py --model pro --think
Show models & pricing
uv run {baseDir}/scripts/models.py
Model Selection Guide
Use Flash when:
- Everyday Q&A and explanations
- Writing, editing, translation
- Code generation and review
- Summarization and classification
- Cost is a priority
Use Pro when:
- Multi-step math or logic problems
- Complex debugging or architecture decisions
- Deep research and analysis
- You want to see the reasoning process (--think)
Tips
- Thinking mode (
--think) streams the internal reasoning before the final answer — useful for hard problems and to verify correctness - System prompt:
--system "You are a concise assistant"to set tone - No streaming:
--no-streamfor cleaner output in scripts - DeepSeek's API is OpenAI-compatible — any OpenAI SDK works with
base_url="https://api.deepseek.com/v1"