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

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-stream for cleaner output in scripts
• DeepSeek's API is OpenAI-compatible — any OpenAI SDK works with base_url="https://api.deepseek.com/v1"