DeepSeek

Getting Started

1. Try DeepSeek online at chat.deepseek.com — free tier available with generous limits.
2. For API access, sign up at platform.deepseek.com — uses OpenAI-compatible endpoints, so just swap your base URL.
3. Run locally with Ollama: ollama run deepseek-r1 for reasoning or ollama run deepseek-v3 for general use.
4. Toggle “DeepThink” mode in the chat for complex math, coding, and logic problems that benefit from extended chain-of-thought reasoning.

Model Lineup

• DeepSeek-R1 — The reasoning model that shook the industry. Transparent chain-of-thought that rivals o1/o3 on AIME, MATH, and coding benchmarks. Shows its thinking process step by step.
• DeepSeek-V3 — A 671B parameter MoE model (37B active) that matches GPT-4 class performance at a tiny fraction of the training cost (~$5.5M vs hundreds of millions).
• DeepSeek-Coder-V2 — Specialized coding model with strong performance across languages and frameworks.

Key Features

• Industry-disrupting cost efficiency — V3 was trained for ~$5.5M, proving frontier AI doesn’t require billions in compute. API pricing is among the cheapest available.
• Transparent reasoning (R1) — Unlike black-box models, R1 shows every step of its chain-of-thought, making it easy to verify and debug its logic.
• Fully open-source weights — Released under MIT license. Download, self-host, fine-tune, and deploy without restrictions.
• Mixture-of-Experts (MoE) — Only activates ~37B of 671B parameters per query, slashing inference costs while maintaining quality.
• OpenAI-compatible API — Drop-in replacement for OpenAI’s API. Change one line of code (the base URL) and you’re running DeepSeek.
• Distilled variants — Smaller versions (1.5B, 7B, 14B, 32B, 70B) distilled from R1, runnable on consumer hardware via Ollama or vLLM.