Is Cursor using the full power of DeepSeek R1?

I tried the same prompt multiple times in:

  1. Cursor Chat
  2. DeepSeek (DS) official website

On DS official website, R1 thought with 2458 words on average.
In Cursor, R1 thought with 523 words on average.

:exclamation:This is only 21% of the thinking done on DS website!

I tried another prompt and the results are similar.

Any thoughts?? Can we get R1 in its full power please?

15 Likes

Can you compare the output rather than thinking?

Less thinking isn’t necessarily bad, maybe cursor is sending optimized prompt to get your output quicker

I’d say DS website output was better because Cursor Chat output ignored an important detail.

2 Likes

most likely restrictions due to high token count

I was about to create a post asking the same.

Is cursor using DeepSeek R1 671B? Or one of the distill versions? because the difference between one of the distill version and the original version is huge.

4 Likes

Yeah, I’m getting the same here (example prompt included): Deepseek-R1 in Cursor seems degraded compared to chat.deepseek.com

I posted it before I saw this question.

2 Likes

According to this comment, it is not a distilled model: Which version of Deepseek R1 does Cursor use? - #2 by deanrie

See explanation from normalnormie here: Deepseek-R1 in Cursor seems degraded compared to chat.deepseek.com - #7 by normalnormie

Cursor team, please look into this issue. Multiple people are complaining about the same. @danperks @deanrie

It feels like Cursor is using another model! Try it for yourself: use the same prompt in Cursor and on DS website. The difference is huge! We want to use R1’s full power. And an essential part of that power comes from longer thinking process.

Some model providers might use the 671B model but quantize it (different from distillation) to use less resources or produce faster responses, but it reduces quality.

2 Likes

Yeah, why isn’t cursor integrating with DeepSeek R1? Claude is clunky and not as good.

yes please cursor team apply some time and take a look into that

Yes, I noticed this the first time I used DeepSeek in Cursor. It stopped when it was in the middle of a file.

Same! deepseek r1 is my new favorite model, full support for the 685B version would be great as I did notice some differences from the model exposed from Cursor to chat.deepseek.com, the quality seems better on the website.

If someone from Cursor could chime in to give some clarity on the matter it’d be awesome!

1 Like

Can someone from cursor team take a look at this please!

Given IP worries, I am concerned about hitting Chinese end points, is Cursor hosting the model itself or a Deepseek API?

they host it in US. Thats why it can be 2-3 times as slow

1 Like

Hey, yes, these models are hosted by the provider Fireworks.

Please use the full-power version of R1. Even if you raise the service fee.

Higher work efficiency can help me earn back more money.

The response by R1 on cursor feels limited when compared to web for some reason T.T

Not sure if not hallucination, but R1 told me he is instructed to “Always provide brief explanation”. That “brief” might cause less thinking and thus worse results. It is only a speculation, but I believe similar wording has negative effect on performance on Perplexity as well. R1 seems to take some user prompts on Perplexity very literally, eg my prompt for Sonnet which was designed so URLs will be outputted as raw URLs, with R1 it forces the model to respond only with list of URLs. So from this I would guess R1 might need more special system prompt tailored just for R1.

2 Likes