DeepSeek
DeepSeek's free chat app and open-weight models: 1M-token context, DeepThink reasoning, vision, and one of the cheapest APIs. A free mini-course from basics to power moves.
What it nails
- Free app with 1M-token context
- Open weights under the MIT licence
- Very low API prices
Watch out for
- Data is stored in China
DeepSeek is a Chinese AI lab whose open models made headlines in January 2025 with the R1 reasoning model. Today its free chat app (chat.deepseek.com and mobile) runs the V4 family: V4-Pro, with 1.6 trillion parameters, and V4.1-Flash (10 September 2026), which adds "native visual understanding". "1M context is now the default across all official DeepSeek services."
Start here: the basics
- Sign in at chat.deepseek.com or in the app with email, Google or Apple ID.
- Pick a mode. Instant answers fast; Expert uses V4-Pro for harder work.
- Switch on DeepThink for problems that need step-by-step reasoning, and Search for current facts.
- Upload files: DeepSeek extracts their text and can read a huge amount at once.
How DeepSeek likes to be prompted
DeepSeek's own advice for its reasoning models: put everything in your message, and for maths ask for step-by-step reasoning with a boxed final answer.
You are a patient physics tutor. Explain why the sky is blue to a 12-year-old in under 120 words, then give one simple experiment they can do at home.
Example by DeepSeek · source ↗ · A teaching prompt. DeepSeek advises: "Avoid adding a system prompt; all instructions should be contained within the user prompt."
A train leaves at 9:40 and arrives at 13:05, stopping twice for 12 minutes each. What was its average moving time per leg if the trip has 3 legs? Please reason step by step, and put your final answer within \boxed{}.Example by DeepSeek · source ↗ · The last sentence is DeepSeek's own recommended wording for maths problems.
I am attaching our company handbook (all 40 files). Build me a one-page table of every rule about remote work, with the file name and page for each, and flag any rules that contradict each other.
A teaching prompt. 1M tokens of context is the default across DeepSeek's official services.
Models and prices
| Model | What it is | API price per 1M tokens (off-peak / peak) |
|---|---|---|
| V4.1-Flash | 552B mixture-of-experts, vision, thinking, 1M context | Input $0.15 / $0.30 · Output $0.60 / $1.20 |
| V4-Pro | 1.6T parameters, 49B active, 1M context | Input $0.66 / $1.32 · Output $1.98 / $3.96 |
The chat app was launched as "100% FREE - No ads, no in-app purchases". On the API, off-peak rates are "half of the peak rates", and cached input costs a fraction of a cent.
Power moves
1. Watch it think
With DeepThink on, DeepSeek shows its reasoning before the answer. Read it: it tells you where the model was unsure.
2. Compare models side by side
Generate an SVG of a pelican riding a bicycle
Example by Simon Willison's Weblog · source ↗ · A famous informal test. Made with DeepSeek-V4-Flash.
Generate an SVG of a pelican riding a bicycle
Example by Simon Willison's Weblog · source ↗ · Made with DeepSeek-V4-Pro. Bigger is not always prettier.
Generate an SVG of a pelican riding a bicycle
Example by Simon Willison's Weblog · source ↗ · Made with DeepSeek-R1-0528.
3. Run it yourself
DeepSeek publishes its weights under the MIT licence on Hugging Face, with ready commands for vLLM, SGLang and Docker. Full V4.1-Flash needs server-class GPUs (about 3 A100 80GB cards at 8-bit, per DataCamp); small distilled versions run on a laptop, with much weaker results.
Generate an SVG of a pelican riding a bicycle
Example by Simon Willison's Weblog · source ↗ · Made with the 8B distilled version run locally with Ollama. Small local models are much weaker.
4. Build on the cheap API
The API "Supports OpenAI ChatCompletions & Anthropic APIs", so most apps built for those can switch to DeepSeek by changing the address and key.
DeepSeek's privacy policy says it will "collect, process and store your Personal Data in People's Republic of China", and chats may be used for training unless you opt out. Several governments restricted it on official devices. Tests by Euronews in 2025 found it avoided politically sensitive topics about China. Do not paste confidential data.
FAQ
Is DeepSeek free?
The chat app is free. The API is pay-as-you-go and among the cheapest available.
What is DeepThink?
A switch that makes DeepSeek reason step by step before answering, for maths, logic and planning.
What is the difference between Expert and Instant?
Instant answers quickly. Expert runs the larger V4-Pro model for harder tasks.
Can I download DeepSeek?
Yes. The weights are open under the MIT licence on Hugging Face, but the full models need powerful GPUs.
Is DeepSeek safe to use?
Data is stored in China and may be used for training unless you opt out. Keep private or business-sensitive information out of it.
Sources: DeepSeek news · V4.1-Flash announcement · API pricing · DeepSeek-R1 usage tips · Privacy policy