HomeFull speed ahead. Keep track of advancements in every frontier

TheTechno Optimist.

DeepSeek-V seriesDeepSeek—DeepSeek-V seriesDeepSeek—DeepSeek-V seriesDeepSeek—DeepSeek-V seriesDeepSeek—

DeepSeek

Open-weight models, V4, R1 reasoning

CNprivateest. 2023deepseek.com (opens deepseek.com)Checked 25 Sep

The brief

Analyst view
  • R1 (January 2025) showed reasoning can emerge from reward-based training with few examples. The paper appeared in Nature.
  • V4 arrived in 2026: a preview in April, then V4-Pro (1.6T parameters, 49B active) in August.
  • Works under US export limits, training on restricted NVIDIA chips and adapting to Chinese chips.
Programs

DeepSeek, two programs

Updated 30d ago
DeepSeek-V seriesDeepSeek's main models. V3 (671B parameters, 37B active) used low-precision maths to train cheaply. V3.2 added a faster attention method. V4 grows to 1.6T parameters.ScalingUpdated 30d agoView program →DeepSeek-R1An open-weight reasoning model trained mostly by rewarding correct maths and code answers. Its R1-Zero version learned to think step by step and check itself without worked examples.DeployUpdated 17 Sep 2025View program →
Research · Proto · Pilot · Deploy · Scaling

Peers

Open-weight models · 4

Company & funding

HQ
CN
Founded
2023

People

Key people
  • Liang WenfengFounder and CEO
Primer