Present · Aug 2025 – today

Hybrid thinking and a million-token context window

V3.1 collapses thinking and non-thinking modes into a single model with native tool calls. Eight months later, V4 Preview pushes the context window to 1 million tokens — Pro at 1.6T parameters, 49B active, API output priced at $3.48 per million tokens, one tenth of OpenAI's equivalent. Native compatibility with Huawei's Ascend 950, open weights as always. The next chapter is being written.

Events in this era

August 21, 2025Read full →
Hybrid Thinking: a first step into the agent era
August 21, 2025: DeepSeek-V3.1 — 128K context, 671B parameters, introducing hybrid thinking. A single model supports both thinking and non-thinking modes without switching, with integrated tool calls. Liang Wenfeng: 'This is our first step toward the agent era.'
April 24, 2026Read full →
V4 Preview: million-token context, dancing with hardware
April 24, 2026: DeepSeek V4 Preview — Pro (1.6T parameters, 49B active) and Flash (284B / 13B). Context window jumped to 1 million tokens via Hybrid Attention. V4-Pro at $3.48 per million output tokens vs OpenAI's $30. Native compatibility with Huawei's Ascend 950. All weights open.