DeepSeek-V3.1

Claims and evidence

• Released 21 August 2025 ● Primary source[1]page header 2025/08/21 [3]Date: 2025-08-21

• Status AvailableWeights are still public on DeepSeek's Hugging Face organisation under the MIT License. In the DeepSeek API, deepseek-chat and deepseek-reasoner moved to DeepSeek-V3.1-Terminus on 2025-09-22. ● Primary source[2]License [3]Date: 2025-09-22
• Replaced by DeepSeek-V3.1-TerminusIn the DeepSeek API only: both API model names were upgraded to DeepSeek-V3.1-Terminus. The weights stay available. ● Primary source[3]Date: 2025-09-22

• Successor of DeepSeek-V3-0324deepseek-chat served DeepSeek-V3-0324 from 2025-03-24 and was upgraded to DeepSeek-V3.1 on 2025-08-21; the model card calls V3.1 an upgrade over the previous version and compares it with DeepSeek V3 0324. ● Primary source[3]Date: 2025-08-21; Date: 2025-03-24 [2]Introduction; Evaluation
• Derived from (continued pretraining) DeepSeek-V3DeepSeek-V3.1-Base was built on the original V3 base checkpoint through a two-phase long-context extension; DeepSeek-V3.1 was post-trained on top of it. ● Primary source[2]Introduction [1]Model Update

• Change · Reasoning One hybrid model serves both a thinking and a non-thinking mode; in the API these modes had been served by separate models, DeepSeek-V3-0324 and DeepSeek-R1-0528.Compared with DeepSeek-V3-0324 ● Primary source[3]Date: 2025-08-21 [2]Introduction
• Change · Tool use DeepSeek says post-training improved tool use and performance on agent tasks.Compared with DeepSeek-V3-0324 ● Primary source[2]Introduction
• Change · Efficiency DeepSeek says the thinking mode gives answers of comparable quality to DeepSeek-R1-0528 while responding more quickly.Compared with DeepSeek-V3-0324 ● Primary source[2]Introduction
• Change · Training data The base model received 840B tokens of continued pre-training on top of DeepSeek-V3 to extend its context length, in a 32K phase and a 128K phase.Compared with DeepSeek-V3 ● Primary source[1]Model Update [2]Introduction

• Feature Reasoning modeOne hybrid model with a thinking and a non-thinking mode; in the API deepseek-reasoner gave the thinking mode and deepseek-chat the non-thinking mode. ● Primary source[1] [2]Introduction
• Feature Function callingTool calls in non-thinking mode per the model card; strict function calling in the Beta API at launch. ● Primary source[1]API Update [2]ToolCall
• Open weights YesMIT License for the repository and the weights. ● Primary source[1]Model Update [2]License
• Context window 128K tokens ● Primary source[2]Model Downloads [1]API Update
• Parameters 671B total, 37B activatedMixture-of-experts model; the model card says the model structure is the same as DeepSeek-V3. ● Primary source[2]Model Downloads
• Access API, consumer app, Open-weights download ● Primary source[1]

Lineage

Variants

Related AI Radar coverage

AI Radar coverage starts in June 2026; no coverage linked yet.

All model releases from DeepSeek on AI Radar →