there seems to be some sort of a token or time limit imposed on the generation process such that it cuts off R1 before it can finish its thoughts and offer code diffs. this behavior is especially notable when you ask rather complicated questions.
Thanks for reporting this. It’s possible that we have a maximum response time for our AIs that doesn’t allow R1 to respond fully. I’ll report this to the team to see if we can get this sorted!
4 Likes
Yes please, it does happen regularly when there is perhaps a bit more output/thinking going on.
2 Likes
I am observing the same experience with deepseek-R1 in Cursor chat
2 Likes
Yep, noticed this.
Some people are doing more than a nodejs project that really needs to make use of “thinking” time…
2 Likes
I agree, it has happened to me when I tried to get him to analyze something quite complex, he really took his time and was doing a very good analysis but at a certain point he got “stopped”.
2 Likes
Happened wth me as well. However sending continue as the follow up message usually seems to work but it is not ideal at all as it eats up fast requests (I think).
sadly this is still a problem…
1 Like
are you guys using Composer with R1 on v0.45 ?
yep, i’m personally on version 0.45.3, and the issue still persists. mostly likely a limit that is set on the backend tbh.
Seems to be a timeout since the cutoff is pretty consistent
1 Like
agreed, looks like it. it’s making it borderline unusable tho with any moderately complex prompt.
1 Like
Thanks for the reports on this!
We’re aware that a timeout on our end is causing R1 to be cancelled before it is finished, but we are hoping to increase this in the next day or so, to allow R1 the time to complete its output.
We are also working on improving the speed of both v3 and R1 in Composer, but this could take longer to get our infrastructure where it needs to be to do so.
6 Likes
this is great to hear, thank you! 
Do we have an ETA for when this will be fixed? I’m repeatedly running into this issue recently when using R1.
Cursor is cutting R1 is not the only problem. R1 is painfully slow in Cursor when compared to other providers. Please fix these two issues in tandem.
Letting it think is not a good enough solution, make it faster too. I really do not want to switch to another editor.
I’ve noticed the same thing, maybe this is due to limitations of fireworks.ai.
There are quite a few new alternatives to that provider coming out, surely soon we will have cheaper options, with more context, and token/s rate available.
Even new models:
- GitHub - huggingface/open-r1: Fully open reproduction of DeepSeek-R1
- DeepSeek R1 (nitro) - API, Providers, Stats | OpenRouter
Everything seems to be developing very fast and people’s demand and expectations are also growing (the latter perhaps too fast), I guess the very existence of this paradigm shift makes even the development faster and that’s why we see things being implemented in Aider like the architect mode / polyglot AI concept so quickly.
1 Like