As an amateur coder, who hasn’t used any paid LLM before, using the DeepSeek R1 model is nothing short of amazing! I felt like I just needed to make a comment about it.
I mainly use cursor for fun, to code some simple game demos, having had quite some success with Claude 3.5 Sonnet. However, when things start getting more complex it has a tendency to introduce bugs, and rearranges code so it breaks things.
I recently gave up on a feature that Claude was having issues with (no matter how much prompt engineering I did), and tried it with the R1 model instead. Reading the thinking output, it understood exactly what I requested, and then made a small concise addition of code that worked on the first try! I’m completely floored by how good it is. This opens up so many doors for an amateur like myself. Thanks Cursor team for implementing this model so quickly!
R1 is the first model that I can say I prefer to use over sonnet 3.5. I don’t really care what benchmark are saying, but in actual usage, R1 feel much stronger. It’s my new favorite.
Yesterday I was switching back to sonnet to see the output with the same prompt and context, and it simply doesn’t compare. I was doing mostly planning, but I will see how good it is at actually writing code, but so far it looks super promising.
Awesome to hear. Thank you for the feedback. For non-reasoning models, starting with an explicit planning step (e.g. “Let’s work this out in a step by step way to be sure we have the right answer. Don’t be lazy.”) can often help in complex, multi-step workflows.
absolutely agree. Seeing how he reasons and thinks about the problem you gave is absolutely mind-blowing. it looks like we have a new king on the throne.
deepseek-chat was amazing for me. After 20 years as an amateur coder at best, I graduated to having more than amateur skill with the chat. What’s the model and api base url pair to use in cursor for R1???
I tested this model yesterday too, but since it’s still a bit slow, I used Sonnet for error correction, and they work great together. I never thought Sonnet would become a secondary model for me. I also think DeepSeek v3 can be used for error correction. Overall, I’m testing, but I’m very impressed so far.
Hi, have you tried r1 on a big or atleast mid scale sized project. how does the quality of output compare vs claude 3.5 . Also for front end development especially UI , claude is very good, have you tested to see if r1 or v3 works as good? thanks
Yes, I tested it on a middle project, mainly on the backend for now, but I will test it for the frontend as well. As for the output quality, it shows significantly better results compared to Sonnet.
I’m trying to use DeepSeek R1 and V3 with Composer (Normal mode), but it’s not editing the files I request. It gets lost and generates several separate blocks… Is Composer working for you guys, or is it just an issue here?
Thanks Cursor for providing Deepseek R1 model amazing model with most code that this model generate think and then generate rather than just throw the certain 100 lines of random code amazing
Hm. So, I love R1. It is really, really a nice and affordable LLM. But I don’t know about its general coding performance. Especially with Cursor, R1 is really slow. Yes, it does its job and doesn’t provide any bugs. But it just takes too much time for me. Sonnet 3.5 is still the king – it’s fast and smart. For bigger problems or analyses, I might switch to R1. How do you guys feel about it?
Yes, I agree. After 24 hours of use, my opinion is more sobering.
While R1 is good for some tasks, I find Claude Sonnet more reliable when building new features. Maybe it’s because Sonnet has Agent mode, but it’s more consistent in nailing functionality.
R1 seems good at reasoning out deeper compile-time errors. Using them together based on their strengths feels pretty good. But in my opinion, Sonnet is still king of code.
Hey Cursor, I know you are all working on it in the background since multi-modal behind the scenes is your jam (I see it very clearly in my billing) but
