Is there any way that you integrate the DeepSeek V3 API so we can use it like GPT etc…?
Should we vote for this?
7 Likes
they are testing deepseek if it meets their requirements they will add it
4 Likes
In both my usage and third-party testing, DeepSeek seems to underperform the latest ChatGPT 4.0 model (gpt-4o-2024-11-20) and Claude’s updated Sonnet (claude-3-5-sonnet-20241022)—albeit by a very small amount. It also underperforms the reasoning models, although that’s not a fair comparison from a latency standpoint.
I think it would be great to put the contours in place such that we can use open-source models within both the Chat and Composer windows. Implementing DeepSeek would serve as a means to an end to getting this done.
That said, I don’t find the DeepSeek model particularly alluring when compared against Claude 3.5 or GPT 4.0 from a purely speed or intelligence standpoint.
Don’t get me wrong—it’s a fantastic model, and I’m very happy these open-source models are basically on par with proprietary models.
But I don’t think implementing DeepSeek into Cursor will result in better outcomes when it comes to interacting or coding with LLMs.
1 Like
It can be slightly worse performing than claude and chatgpt (IMO it is more than slightly worse especially for longer tasks), but it is so much cheaper than both of those things that it is very much worth it to use it on smallish tasks and move to claude only if it is not doing so well. Because cursor doesn’t have support for this, I use cline / aider with deepseek. This reduces my dependency on claude by a significant amount. It would be fantastic to be able to use this within cursor itself without having to switch tools.
Yes I kow you can use open ai compatible endpoints and stuff, but that means i have to disable all the other models and I don’t want to do that.
3 Likes
do that, but hey, add Azure’s o1 too finally
I don’t know what type of tests you were running, but in a real development situation, DeepSeek crushes the competition. In a real use case with a massive database of knowledge to sort through, DeepSeek costs less and so far shows ZERO signs of hallucinations. But for coding… deepseek-coder destroys the competition and deepseek-chat performs better than any openai gpt option. DESTROYS
It’s easy.
Just use deepseek-coder as the model and the Base URL: https://apiDOTdeepseekDOTcom/
Can you show some example prompts where you’re seeing DeepSeek significantly outperform the latest Claude Sonnet model?
Those are benchmarks from Deepseek’s own paper. I was looking for practical examples.
(and even that shows Claude 3.5 Sonnet outperforming in 3/6 tests).
Lol… Pick any open source project. Purposely create an error in it. Type “Fix the error in the code” in the chat prompt and hit Shift+Enter… Use Claude and watch. Use deepseek and watch. ALSO>>>>. Watch the costs. Make sure you have a large .cursorrules file so your code is top notch 
Hey all, DeepSeek v3 is now available to test in the editor!
You will have to go in and enable it in your model settings first, but it is included as a non-premium model, so give it a try!
We’ve added this as a bit of an experimental model, so please note it could be removed at any time.
1 Like
Any ETA for DeepSeek R1? It’s showing great promise and could definitely take some heat off Sonnet.
2 Likes
Yes indeed, but i think it`s enough
Yeah, these models require more work on our end to get images to work, but I’ll mention it with the team to see if we can get that added!
1 Like
ty, keep us updated!
I believe DeepSeek v3 does not have built-in support for images, hence the error inside Cursor.
While their web UI has support for images, we believe an intermediate model is used to extract the information from any uploaded images, which is then passed to DeepSeek v3.
The two are on par with each other. However, Deepseek is cheaper than Claude.