Skip to content

[FEATURE]: DeepSeek V3.2 Support #4796

@vladnosiv

Description

@vladnosiv

Feature request

DeepSeek V3.2 Support

Describe the problem you're encountering

The new model has a new chat template, as well as changes in tool parsing: the model is now able to call tools during reasoning.

DeepSeek posted the preprocessing and parsing code (can be used as reference for implementation): https://huggingface.co/deepseek-ai/DeepSeek-V3.2/blob/main/encoding/encoding_dsv32.py

It is important that the colleagues from DeepSeek write that due to the complexity of the new template, they intentionally didn't add the jinja template and recommend using their manual rendering. Currently, Dynamo supports only jinja templates.

Describe alternatives you've tried

No response

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions