DeepSeek has announced the final iteration of its V2 series models with the release of DeepSeek-V2.5-1210. This version marks the culmination of the V2 series before the launch of an entirely new base model series.
The DeepSeek-V2.5-1210 model features significant improvements across various abilities through comprehensive post-training, including mathematics, coding, writing, and role-playing.
In line with their commitment to open-source, the weights for the new model are available on Hugging Face.
https://huggingface.co/deepseek-ai/DeepSeek-V2.5-1210
The DeepSeek-V2.5-1210 introduces the added functionality of web search, currently available on the web interface at https://chat.deepseek.com/. Activate the "Web Search" option within the input box to enable this feature. (Please note that the API does not currently support this functionality).
In "Web Search" mode, the model analyzes vast amounts of online data to provide comprehensive, accurate, and personalized responses. For complex queries, it extracts multiple keywords to conduct parallel searches, delivering diverse results more efficiently.
DeepSeek-V2.5-1210 marks the end of the DeepSeek V2 series. Launched in May, the V2 series has undergone five iterations, driven by user support and feedback.
DeepSeek is now focused on building the next-generation base model, DeepSeek V3. Stay tuned for updates!
Are you interested in trying out DeepSeek's API? Check out our Quick Start Guide to start building awesome applications.
You can find more information about models and pricing here.
