DeepSeek Introduces Vision (chat.deepseek.com)

Article summary

The article announces the introduction of Vision by DeepSeek, but the details of the article are not available. The discussion reveals that Vision is not yet available in the API and there is no official announcement about its release. Users are waiting for the vision support in the API to subscribe to the service. The discussion also touches on the model's language preferences and reasoning processes.

Main themes

• DeepSeek Vision
• API availability
• Language preferences
• Model reasoning
• Chinese language support
• Model updates

What commenters say

• The model's reasoning traces are generated in human languages, not an internal 'alien language'.
• Some models may produce reasoning that is not always understandable for humans, but this is not the case for DeepSeek.
• The quality of the model's output scales with the amount of 'thinking' tokens it generates before responding, despite potential contradictions in the thinking process.
• The model's language preference for Chinese may be due to better datasets and user data for training, leading to a 'silent update' with increased Chinese responses.
• The use of images instead of raw text can be a more efficient way of processing and reasoning for models.
• The model's ability to think in different languages, including Classical Chinese, can be beneficial for reducing token usage and improving performance.
• The inconsistency in the model's language output, sometimes switching to Chinese and forgetting to switch back to English, is a notable issue that has been observed by users.
• The model's behavior of responding in Chinese may be a result of distillation, where the model is influenced by user input and adapts its language preference accordingly.