DeepSeek R1 has landed on Cursor Composer, so I tried it by adding Supabase realtime to my app! If you're wondering, they are using a version of DeepSeek hosted on US servers. As much as I love the hype on this model, I find that using Claude Sonnet produces... better results.show more

Yeah Sonnet is still way ahead in terms of coding! But reasoning traces from R1 can be used in other models (feature), and that's exactly what guys from DeepClaude did - they use R1 to create a reasoning trace and then use it to improve output from Sonnet - mindblowing!

Learn Japanese through "mild immersion" in a game Like this? Repost and follow us!

Personally I think where things are going to really benefit from Deepseek is when people decide to use it for reasoning only and then pass that reasoning to Claude to implement

Use R1 to create a plan for working on a feature. Have it break that down into chunks it can assign to jr developers. Then feed that plan into Composer using Claude 3.5. Works really well. R1 is the architect, Claude is the Staff engineer

I'd like it in Cursor if I could use R1 offline when I can't use Sonnet (like the flight I'm imminently departing on where the Wi-Fi is broken)... PS I thought it was funny when @cursor_ai announced R1 support, but said that they're still leaving Sonnet as default due to better performance

You're using it wrong. You should use R1 in chat mode to analyze a problem and provide step-by-step instructions, then copy the entire output (including the thinking) over to the agent composer with Claude 3.5 Sonnet. I've found the results astonishing.

Have you seen that you can now create and manage Supabase projects now with Cursor Composer thanks to MCP support? That’s the game changing feature for me

DeepSeek R1 is slow, inaccurate, and often confusing itself. Buying more $NVDA everyday!

@supabase Agreed , 5 days of testing, and I still don’t get the hype.

@supabase Yep, I’ve still yet to find a model that outperforms Claude despite what benchmarks may have us believe

it's amazing that still no model can beat sonnet