DeepSeek Provider

DeepSeek is a Chinese AI lab producing high-quality open-weight models with strong coding performance at very affordable prices. DeepSeek V3 and R1 consistently rank among the top models on coding benchmarks.

Get an API Key

Go to platform.deepseek.com
Create an account or sign in
Navigate to API Keys
Create a new key and copy it

Configure VibeCody

Option 1: Environment variable (recommended)

export DEEPSEEK_API_KEY="sk-..."
vibecli --provider deepseek

Option 2: Config file (~/.vibecli/config.toml)

[deepseek]
enabled = true
api_key = "sk-..."
model = "deepseek-chat"

Model Selection

Model	API Name	Strengths	Best for
DeepSeek V3	`deepseek-chat`	General-purpose, strong coding	Code generation, debugging, refactoring
DeepSeek R1	`deepseek-reasoner`	Chain-of-thought reasoning	Complex logic, architecture decisions, hard bugs

Default: deepseek-chat (V3)

Override from the CLI:

vibecli --provider deepseek --model deepseek-reasoner

When to use R1 (Reasoner)

DeepSeek R1 uses chain-of-thought reasoning similar to OpenAI’s o-series. Use it when you need:

Multi-step debugging of complex issues
Algorithm design and optimization
System architecture planning
Mathematical or logical reasoning in code

For everyday coding tasks, deepseek-chat (V3) is faster and more cost-effective.

Pricing

DeepSeek is one of the most affordable providers available:

Model	Input (per 1M tokens)	Output (per 1M tokens)
DeepSeek V3 (chat)	$0.27	$1.10
DeepSeek R1 (reasoner)	$0.55	$2.19

This makes DeepSeek roughly 5-10x cheaper than GPT-4o and 10-50x cheaper than Claude Opus for equivalent tasks.

Cache hits are even cheaper – DeepSeek caches common prefixes automatically.

Best For

DeepSeek excels at:

Code generation – strong performance across Python, JavaScript, Rust, Go, and more
Debugging – identifies bugs and suggests fixes with high accuracy
Code review – catches issues that other models miss, at a fraction of the cost
Refactoring – understands complex codebases and suggests clean improvements
Batch workloads – very affordable for high-volume agent tasks

Verify Connection

vibecli --provider deepseek -c "Write a Rust function to merge two sorted arrays"

Running DeepSeek Locally

DeepSeek models are open-weight. You can run them locally via Ollama:

ollama pull deepseek-coder-v2:16b
vibecli --provider ollama --model deepseek-coder-v2:16b

This gives you the DeepSeek model quality with zero API costs and full privacy.

Troubleshooting

Invalid API key

Error: 401 Unauthorized

Check the key at platform.deepseek.com
Confirm the env var is set: echo $DEEPSEEK_API_KEY

Rate limited

Error: 429 Too Many Requests

DeepSeek has rate limits based on your account tier
Wait briefly and retry
Contact DeepSeek support to increase limits for production use

Slow responses with R1

The reasoner model (deepseek-reasoner) performs chain-of-thought reasoning internally, which can take 15-60 seconds for complex prompts. This is expected behavior – the model is “thinking” before responding.

Connection issues

Error: Connection timed out

DeepSeek servers are primarily hosted in China
Some regions may experience higher latency
Consider using a proxy if you have consistent connectivity issues
Alternatively, run DeepSeek locally via Ollama for zero-latency inference

Provider: DeepSeek