O3-mini supported the second it's out. Why isn't Deepseek-R1 supported in Composer's Agent mode?

Discussion
1

Why isn’t Deepseek-R1 supported in Composer’s Agent mode?

P.S: I know it’s supported in Chat, but I don’t use it for chat; I need R1 in the agent.

Does Cursor intend to add it to Composer anytime soon? Any timeline?

EDIT:

According to the Aider leaderboard (Aider LLM Leaderboards | aider):

image
image1144×1031 68.3 KB

DeepSeek-R1 has a 96.9% success rate at using the correct edit format in aider’s benchmark. That’s literally testing if a model follows prompts properly for returning the required format. For comparison:

  • Claude-3.5-Sonnet: 99.6% (better)
  • o3-mini (high): 93.3% (worse)
  • DeepSeek-R1: 96.9%

Deepseek-R1 also performs better than o3-medium and Claude Sonnet 3.5 in terms of correct answers…

So if Cursor IDE doesn’t support DeepSeek-R1, it’s not because of “bad prompt following” — the data proves the opposite. Aider’s benchmark specifically measures whether models adhere to the system prompt to edit code without human intervention, and DeepSeek-R1 ranks among the top performers (according to the numbers).

If models with worse prompt-following scores are supported, why isn’t the better-performing DeepSeek-R1 included? “That math ain’t mathing”.

1 Like
2

Maybe because GPT-4o was already supproted in Agent mode, so Cursor have already OpenAI’s models’ integration figured.

They are working on Deepseek also as we have been told.

3

cause deepseek dont have a good tool calling/prompt following and cause of that during the prompt processing things will break

1 Like
4

You may have a point in tool support, but the argument about “prompt following” is not backed by the data (cf. my edit in my original post)

5

As others have suggested, o1-mini was already tuned to work well in Composer, so o3-mini behaved very similarly behind the scenes (although did still have some teething issues!)

DeepSeek behaves entirely different, and requires some more trial and error to get it to perform well with the Agent, but we are actively working on this now and will hopefully be able to add this soon!

2 Likes