DeepSeek
The Chinese open-source AI model that shocked the world. Performance close to Claude and GPT at a fraction of the price — but with question marks around security and data.
What I love
- + Open-source model (DeepSeek V3, R1)
- + Performance comparable to GPT-4o at a fraction of the price
- + Strong at math and reasoning (R1)
- + You can run it locally (Ollama, vLLM)
- + The API is extremely cheap
What bugs me
- − Chinese company — questions around data and censorship
- − Czech is noticeably weaker
- − Censors politically sensitive topics (China)
- − Smaller ecosystem of tools
- − Unstable API (outages)
The elephant in the room
DeepSeek shocked the AI community. An open-source model out of China that competes with GPT-4o and Claude Sonnet on benchmarks? At a fraction of the price? I had to try it.
What's true
Performance
DeepSeek V3 and R1 are legitimately good models. On code and math they reach results close to top-tier models. R1 (reasoning model) beats o1 on some benchmarks.
I tested on real tasks:
- Code (simple): Comparable to Sonnet
- Code (complex multi-file): Behind Claude, but functional
- Math/logic: Surprisingly strong, R1 excels
- Czech: Weak. Noticeably behind Claude and GPT
Price
DeepSeek wins absolutely here. API at $0.27/1M input tokens vs. $3 for Sonnet. A tenfold difference. For batch processing and non-sensitive tasks it's interesting.
Open-source
You can download the model and run it locally. Via Ollama it runs on decent hardware (A100, RTX 4090). For companies with strict data residency requirements, that's key.
Why only 3 stars
Security and data
DeepSeek is a Chinese company. Data processed through their API is subject to Chinese law. For personal projects — OK. For corporate data under NDA — hard no. Local hosting solves this, but requires hardware.
Censorship
Ask about Taiwan, Tiananmen, or Tibet. DeepSeek either refuses to answer or gives a "diplomatic" response. For politically neutral tasks (code, math) it doesn't matter. For general use it's a dealbreaker.
Stability
The API has outages. Sometimes it's slow, sometimes unavailable. For production deployment you need a fallback.
When DeepSeek makes sense
- Batch processing of non-sensitive data (price!)
- Math and reasoning tasks (R1)
- Local hosting on your own hardware
- Experimentation and prototyping
Verdict
DeepSeek is an impressive technical achievement and a legitimate choice for specific use cases. But for daily work in Czech, for sensitive data, for a production API? Claude or GPT. DeepSeek is a fascinating experiment and a strong signal that the AI race is global — but not my daily driver.