Congrats, this is the by far the best open source model! Just a few steps until complete domination (feedback)
This model writes, codes and reasons very well. I was extremly impressed. Not only feels the language natural, it also has a great personality which is something not many LLMs are able to capture and it has far greater creative writing capabilities than competing brands such as Qwen and OpenAI.
Now, the next step in my opinion is to combine everything you have learned and make one model that excels at every task. Have a model that accepts video, audio and text inputs and perhaps even outputs, native omnimodality from the ground up. Have it reason like R1 on request or not in resource constrained environments (even if I specifically tell R1 not to reason, it does it, in a future general purpose model this could be steered via a system prompt). Let it have excellent function calling support and build it on a groundbreaking new architecture such as the recently released paper for the Titans architecture (https://aipapersacademy.com/titans/)
Sincerely, thank you for open sourcing your baby. I'm very excited what the future brings for DeepSeek. Keep up the great work!
how's your view about integrating with google's work titans?
e.g. used in SFT or RL stage or else ideas?