← All models

DeepSeek-V3.2-Exp

Published Publisher DeepSeekdeepseek-chat

Overview

DeepSeek-V3.2-Exp is an experimental large language model released on September 29, 2025. Built on DeepSeek-V3.1-Terminus, it introduces DeepSeek Sparse Attention (DSA) to improve training and inference efficiency for long-context tasks. DeepSeek reported broadly comparable performance to its predecessor across public benchmarks. (deepseek.com)

Capabilities and uses

The model supports text generation and conversational applications, including coding, mathematical reasoning and agent-oriented tasks. It was released for experimentation with efficient long-context processing. At launch, deepseek-chat selected its non-thinking mode; deepseek-reasoner selected its thinking mode. (huggingface.co)

DeepSeek published the model weights and repository under the MIT license, along with GPU kernels and a technical report. (huggingface.co)

Limitations and availability

This is an experimental release: DeepSeek cautioned that broader real-world testing was needed to identify scenarios where performance might be weaker. Outputs should be checked for important uses. The API ID is not a version-pinned identifier: DeepSeek upgraded deepseek-chat to the formal V3.2 model on December 1, 2025, and subsequently announced its discontinuation. It should not be assumed to serve V3.2-Exp today. (api-docs.deepseek.com)

Historical official API pricing

The following are launch-era prices, not a current offer for V3.2-Exp. Each price is per 1 million tokens; input and output are billed separately. The two regional quotations are recorded in their original currencies, not converted into one another.

Official sources