google
DeepSeek v3.2, tools, thinking, and messages
Been reading through the DeepSeek 3.2 paper, and this section stuck out, and reminded me that I have questions about why we use "model"/"user" messages for tool events, generally, and specifically to ADK
Mostly want to start a thread to talk about this. It seems more message types would be useful, but perhaps the models are not good with them because they are not post-trained on the variety (yet)
Very interesting, "model"/"model" messages for tool events is definitely a topic to further explore. Thank you for bringing this specific DeepSeek behavior to our attention!
@baptmont fwiw, I believe they are using the Harmony response format without saying so.
Not sure if everyone is moving towards Harmony at this point or not, curious where Google's head is at on this
https://cookbook.openai.com/articles/openai-harmony