Key facts

Founded
—
HQ
Hangzhou, China
Employees
—
Website
Funding stage
—
LinkedIn
—
Further facts
  • V4 Preview: DeepSeek open-sourced V4 in April 2026, with V4-Pro at 1.6 trillion total and 49 billion active parameters and a 1 million token context.
  • Latest model: Released in September 2026, V4.1-Flash uses a new causal encoder-decoder design with 8 billion active parameters for input and 16 billion for output.
  • Model routing: From 14 September 2026, requests to V4-Pro route to V4.1-Flash until V4.1-Pro launches.
  • Pricing: Off-peak API rates are 50% lower than peak rates.
  • Search API: Search and Fetch APIs opened to all API users in October 2026 at US$0.75 per 1,000 calls per endpoint.

Focus areas

Open-weight models Reasoning models Foundation models Multimodal AI API platforms

About

DeepSeek is a Hangzhou AI lab that builds large language models and publishes many of them as open weights on Hugging Face. Its V4 Preview, open-sourced in April 2026, brought a 1 million token context by default, and V4-Pro reached general availability in August 2026. In September 2026 it released V4.1-Flash, the first model in a new architecture family, and began phasing out V4-Pro. DeepSeek offers its models through its website and app, a developer API and its own Search and Fetch APIs.

How they use AI

DeepSeek trains its own large language models. V4 combines token-wise compression with DeepSeek Sparse Attention to cut compute and memory costs at long context. V4.1-Flash adds native image understanding and a smaller KV cache, and the October 2026 V4.1-Flash-1017 update uses reinforcement learning with self-feedback for end-to-end project building. The API supports OpenAI and Anthropic formats, the OpenAI Responses API and adjustable reasoning effort, and works with coding tools such as DeepSeek Harness, Codex and Claude Code.

Products and services

  • DeepSeek-V4.1-Flash — Current API model with native image understanding, open weights on Hugging Face
  • DeepSeek-V4 — Open-weight V4-Pro and V4-Flash models with a 1 million token context
  • DeepSeek API — Developer API compatible with OpenAI and Anthropic formats, with peak and off-peak pricing
  • Search and Fetch APIs — Web search and page fetch powered by DeepSeek's own search engine
  • DeepSeek Harness — DeepSeek's agent harness with built-in support for its models
  • DeepSeek app and chat — Consumer chat on web and mobile