Open weights

One-line definition

Releasing the trained model file under a permissive license, not just inference code. DeepSeek's V2 / V3 / R1 / V3.1 / V4 weights are open downloads.

Full explainer

Open weights means publishing the trained model file under a license that permits redistribution and commercial use — not merely open-sourcing the inference code or sharing API access. DeepSeek has released every flagship model since Coder (November 2023) under such terms; R1 specifically uses the MIT license, and the V3 / V3.1 / V4 model cards on HuggingFace are public downloads. This is the technical pre-requisite for everything downstream of the model: private deployment, fine-tuning, jurisdictional sovereignty, distillation, and audit. Closed-API releases (OpenAI, Anthropic) cannot offer these. The cultural framing — 'open-sourcing papers loses you nothing' — comes from Liang Wenfeng's 36Kr interview and is the line that most distinguishes DeepSeek's stance from frontier U.S. labs.