DeepSeek-V4.1-Flash-1017: Format chats with its Python encoder

What it does

DeepSeek-V4.1-Flash-1017 is listed as an image-text-to-text model, so it takes text and images in a conversation. This release has no ready-made chat template. Instead it ships a Python reference program that turns a conversation into the exact prompt format the model expects, with test cases for multi-turn conversations, thinking mode, system messages and images mixed into the conversation.

How to use it

Input → action → output: write your messages, adding images if needed, encode them with the supplied encoding code or the deepseek-recipe libraries, run the model by following the inference folder's instructions, then decode the reply. Running the model itself is left to you.

Common use cases

• Continue a multi-turn conversation where one of the messages includes an image, and read the model's reply.
• Turn on thinking mode in a conversation so the model reasons before it answers, using the supplied encoder to format the prompt.