DeepSeek V3.2

by DeepSeek

About the model:

Mixture-of-experts chat model.

Model type:

Text

Model performance class:

Latest

Link to the model:

https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Speciale

Base models:

DeepSeek-V3.1-Base

End model:

DeepSeek-V3.2-Speciale

End model license:

MIT License

About the organisation:

DeepSeek (深度求索), founded in 2023, is a Chinese company dedicated to making AGI a reality.

Link to the organisation:

https://www.deepseek.com/

Model release date:

November 2025

Availability

Base Model Data

Pretraining data not specified or documented besides the claim that it amounts to 14.8T tokens in Chinese and English. Proprietary dataset.

End User Model Data

Post-training data not specified or documented. Said to include '1.5M instances'.

Base Model Weights

Weights made available on HuggingFace.

https://huggingface.co/deepseek-ai/DeepSeek-V3-Base

End User Model Weights

Weights made available on HuggingFace.

https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Speciale

Training Code

Outdated repository provides sparse source code and some examples for inference.

https://github.com/deepseek-ai/DeepSeek-V3/tree/main/inference

Documentation

Code Documentation

Repository is sparsely documented.

https://github.com/deepseek-ai/DeepSeek-V3/

Hardware Architecture

Model architecture described in a fair bit of detail in technical report.

https://arxiv.org/abs/2512.02556

Preprint

Corporate technical report provides info on infrastructure compute, performance, but nothing on pre- or post-training data.

https://arxiv.org/abs/2512.02556

Paper

No peer-reviewed paper found.

Modelcard

Model card on HF provides basic summary but no info on intended, scope, biases, risks and limitations.

https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Speciale

Datasheet

No datasheet provided.

Access

Licenses

MIT License, an OSI-approved license

https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Speciale#license

Last updated 13 March 2026

Is this information not up to date?

Contribute here ->

Supported by the Centre for Language Studies and the Dutch Research Council. Website design & development © 2024 by BSTN. This version of the index generated 09 April 2026, website content last updated 11 March 2026.