hi,
i’m noticing a difference in quality between the responses from the official website and cursor. i seem to be getting better results with R1 when i use the official website compared to cursor, with the same prompts. now this could be due to a variety of reasons, and i don’t want to rush to any conclusions. but it did make me wonder whether cursor is using a quantized version of R1.
i’d appreciate any clarification on this. thank you.
It’s exactly my question as well. I’ve been experiencing the same quality issues. The one received through my own API is far better than the one Cursor uses. As they mentioned earlier, they’re using Fireworks for that. Maybe they don’t use the full version, which significantly downgrades the quality of the output.
Hey, according to this page, this model has 685B parameters, so yes, it’s the full version:
But is it already confirmed that Cursor uses the 671B parameters model from Fireworks? Maybe it uses another Distill R1 version, or adjusts the context more aggressively, it would be good to confirm this officially.
User @danperks was going to do this in another post when he gets that answer.
The team confirmed that the model with 671B parameters is used, not the Distill R1 version.
Is it quantized in any way or is it running in fp8 / bf16?