DeepSeek

DeepSeek is an AI lab whose products and services are owned and operated by Hangzhou DeepSeek Artificial Intelligence Co., Ltd.¹ The oldest repositories on its deepseek-ai Hugging Face account, DeepSeek Coder models, date from October 2023,² and its first language-model paper, "DeepSeek LLM: Scaling Open-Source Language Models with Longtermism," was posted on January 5, 2024.³ Its own release log runs from DeepSeek-V2.5 (September 2024) through DeepSeek-V3 (December 26, 2024), DeepSeek-R1 (January 20, 2025) and DeepSeek-V3.2 (December 1, 2025) to the DeepSeek-V4 preview (April 24, 2026) and DeepSeek-V4.1-Flash (September 10, 2026).⁴ It publishes weights on Hugging Face, with V3.2, V4-Pro and V4.1-Flash under the MIT License,⁵ and sells API access in formats compatible with OpenAI's and Anthropic's APIs.⁶

DeepSeek's newest model leads Most-Used AI Models on OpenRouter by Weekly Tokens: DeepSeek V4.1 Flash processed 40.53 trillion tokens through OpenRouter in the seven days to October 12, 2026, up 43% on the week before, and two earlier V4 Flash versions are also in the top 20.⁷ In Cheapest LLM APIs per 1M Output Tokens, the deepseek-flash model lists $0.30 per 1M input tokens and $1.20 per 1M output tokens at peak hours, deepseek-v4-pro lists $1.32 and $3.96, and off-peak rates are half.⁸ In Chinese Open-Weight Models: Downloads, Derivatives and Token Share, the ATOM Report finds DeepSeek's cumulative Hugging Face downloads passing Mistral's in January 2026 (128.2 million against 124.3 million) and DeepSeek holding 31.1% of open-model tokens on OpenRouter that month,⁹ and Hugging Face reports that DeepSeek-R1 became the most-liked model on the Hub within a year of its release.¹⁰