🧠

DeepSeek Reasoning (R1)

DeepSeek Reasoning is the dedicated "thinking" model line from DeepSeek, a Chinese AI lab — starting with DeepSeek-R1 and continuing through its successors. It ranks #7 in real-world usage share at roughly 23.3%, and made global headlines in early 2025 by matching the reasoning quality of top closed-source models at a small fraction of the training and inference cost. Unlike most frontier reasoning models, its weights are fully open, so anyone can download and self-host it.

Quick facts
Made by: DeepSeek
Access: Open-weight, self-hostable (Ollama, vLLM, Hugging Face) + a low-cost hosted API
Context window: ~64K–128K tokens depending on version
Strengths: Step-by-step chain-of-thought reasoning, math, coding logic, transparent "thinking" traces
Primary use case: Hard reasoning tasks — math proofs, multi-step coding problems, logic puzzles

Jump to: Example Getting started Best for

Example

Run the model fully locally with Ollama, or call DeepSeek's OpenAI-compatible API endpoint directly.

# run locally with Ollama
ollama run deepseek-r1

curl https://api.deepseek.com/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -d '{
    "model": "deepseek-reasoner",
    "messages": [{"role": "user", "content": "A train leaves at 3pm going 60mph. Another leaves at 4pm going 90mph on the same track. When does it catch up?"}]
  }'

Getting started

The fastest path is running it locally with no API key at all — or signing up for DeepSeek's hosted API if you want managed hosting.

# 1. install Ollama, then pull the reasoning model
ollama pull deepseek-r1

# 2. or, for the hosted API — sign up at platform.deepseek.com
# then set your key
export DEEPSEEK_API_KEY="your-key-here"

Best for: Developers who need a model to visibly "show its work" on hard multi-step problems — math, logic, and debugging tricky code — and who want the option to run it entirely on their own hardware with no per-token cost.