# DeepSeek V3.2 reimplementation

**URL:** <https://discuss.huggingface.co/t/deepseek-v3-2-reimplementation/176595>\
**Category:** Show and Tell\
**Created:** [June 6, 2026, 9:10pm UTC](https://discuss.huggingface.co/t/deepseek-v3-2-reimplementation/176595 "2026-06-06T21:10:20Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**Author:** ![JonathanC123](https://avatars.discourse-cdn.com/v4/letter/j/c37758/32.png) [@JonathanC123](https://discuss.huggingface.co/u/JonathanC123)\
**Post date:** [June 6, 2026, 9:10pm UTC](https://discuss.huggingface.co/t/deepseek-v3-2-reimplementation/176595/1 "2026-06-06T21:10:20Z")

</div>

Hi everyone, I saw on huggingface that they had implemented DeepSeek V3 but not DeepSeek V3.2. So as a learning opportunity I decided to implement a small scale deepseek v3.2 reimplementation with some small scale implementations. It can be found here:

> **[GitHub - JonathanColetti/deepseek-v3.2-pytorch: A unofficial PyTorch implementation of DeepSeek...](https://github.com/JonathanColetti/deepseek-v3.2-pytorch)**
>
> A unofficial PyTorch implementation of DeepSeek V3.2 in pytorch

It implements the new DSA (lightning indexer and token selection mechanism (top k). If you have any questions or would like to discuss more with me let me know!