DeepSeek AI

R1 is a reasoning-focused model that performs visible chain-of-thought before answering, comparable to OpenAI o1 on math and coding benchmarks. It is released under MIT-style licensing with full open weights available on Hugging Face, enabling self-hosting and fine-tuning. API pricing is approximately $0.55 per million input tokens, dramatically undercutting closed competitors.

V3 uses a 671-billion parameter mixture-of-experts design with only 37 billion parameters activated per token, delivering frontier-class performance with relatively efficient inference. It powers the default chat experience and excels at general-purpose tasks like writing, summarization, and coding. The model was trained on 14.8 trillion tokens with multi-token prediction objectives.

DeepSeek's API endpoint mirrors OpenAI's request and response schema, so existing applications using openai-python or openai-node can switch by changing the base_url and api_key. This compatibility supports streaming, function calling, JSON mode, and chat completions. Migration typically takes minutes rather than days, removing a major barrier to switching providers.

The chat.deepseek.com web interface and native iOS/Android apps offer unlimited free access to V3 and R1 for consumer use, including web search and file upload. No subscription is required, distinguishing it from ChatGPT Plus and Claude Pro which gate their best models behind paid tiers. The mobile app briefly became the #1 free app on the U.S. App Store in January 2025.

Both DeepSeek-V3 and DeepSeek-R1 model weights are publicly downloadable under permissive licensing terms, including for commercial use. Distilled smaller variants (1.5B to 70B parameters based on Qwen and Llama) are also available for resource-constrained deployments. This openness has driven rapid community adoption and integration into platforms like Ollama, vLLM, and LM Studio.