Model positioning
DeepSeek-V3.1-Base is the base pretrained language model for DeepSeek-V3.1. It was further trained for long-context extension on the original DeepSeek-V3 base model checkpoint; the post-trained DeepSeek-V3.1 model is built on this foundation. Its release was officially announced on August 21, 2025. (deepseek.com)
Capabilities and uses
- The model card lists 671B total parameters, with 37B parameters activated per token, and a context length of 128K tokens. (huggingface.co)
- According to the official announcement, the Base model was trained on approximately 840B tokens in addition to V3 to extend its long-context capabilities. It can be used for text generation and as a foundation for post-trained or customized models. (api-docs.deepseek.com)
- The model weights have been officially released; the model card specifies an MIT license. The license in the model repository governs the specific terms of use. (api-docs.deepseek.com)
Limitations and deployment considerations
This is a base model; the thinking and non-thinking modes, tool calling, or agent evaluation performance of the post-trained DeepSeek-V3.1 should not be taken directly as guarantees for the Base version. The V3.1 series made substantial changes to the tokenizer and chat template, so the updated configuration should be checked during deployment; deploying a large-scale model also requires sufficient computing and memory resources. (huggingface.co)
Pricing
As of November 18, 2026, DeepSeek's official domestic RMB and international USD pricing pages were checked, and no official API input or output pricing specifically for DeepSeek-V3.1-Base was verified. The API modes and pricing adjustments in the announcement cannot be directly treated as pricing for the Base model. (api-docs.deepseek.com)