DeepSeek

Overview

DeepSeek is a Chinese AI research lab founded in 2023 by Liang Wenfeng. The company operates as a subsidiary of High-Flyer Capital Management, a quantitative hedge fund based in Hangzhou, and is entirely self funded from hedge fund profits with zero external venture capital. DeepSeek became a global phenomenon in January 2025 when its R1 reasoning model matched GPT-4 level performance while costing 10 to 50 times less to run. This event, known as the "DeepSeek moment," triggered a sell off in US tech stocks, reignited debates about AI chip export controls, and forced the entire industry to confront the reality that a relatively small Chinese lab could match the output of companies spending tens of billions on compute. The company's flagship model is DeepSeek V3.2, a 671 billion parameter Mixture of Experts (MoE) model with 37 billion active parameters per forward pass. All DeepSeek models are released under the MIT license with full weights, training methodology, and technical papers published openly. API pricing sits at $0.28 per million input tokens and $0.42 per million output tokens, making it roughly 10 to 50 times cheaper than comparable Western frontier models. DeepSeek's product lineup includes DeepSeek V3.2 (the flagship language model), DeepSeek R1 (a reasoning model with R1-0528 as the latest variant), DeepSeek Chat (a free web interface), and the DeepSeek API (a developer platform with extremely low pricing).

What makes them different

Compared to OpenAI and Anthropic, DeepSeek achieves comparable model quality at dramatically lower cost for both training and inference. Where OpenAI and Anthropic charge $5 to $15+ per million input tokens for frontier models, DeepSeek charges $0.28. The company also publishes full training details and model weights, while OpenAI and Anthropic keep their models proprietary. The trade off is that DeepSeek's consumer product (the chat interface) is far less polished than ChatGPT or Claude, with fewer features and a simpler interface. Compared to Meta's Llama, both release open weights, but DeepSeek publishes significantly more detail about its training process and methodology. DeepSeek's models also tend to outperform Llama on reasoning and coding benchmarks, though Meta has a substantial ecosystem and distribution advantage through partnerships with major cloud providers. Compared to Google's Gemini, Google has the broadest AI product suite and deepest integration with consumer services. DeepSeek competes purely on model quality and cost efficiency with no consumer ecosystem, no search engine, and no hardware division. Unique strengths include being the most cost efficient frontier AI models in the world, fully open source with MIT licensing and published training methodology, zero reliance on external funding or investor pressure, pioneering Mixture of Experts (MoE) and Multi-head Latent Attention (MLA) innovations that the broader industry has adopted, and proving that frontier AI is achievable without massive Western scale compute budgets.

Their tools

DeepSeek

chatbot

deepseek

Frontier AI models that are completely free to chat with, the cheapest to build on, and fully open source under the MIT license.

FreeView details →

Subscription plans

Links