[127]

Strange. I'm switching from Codex and Claude to Pi with Qwen3.6 27B local and Deepseek V4 Flash which is dirt cheap but powerful.

[Kuyawa]

I've built five apps in the last month using DeepSeek, spending less than $1 in total. I am totally in love with DeepSeek and my wife knows it :)

[satvikpendem]

What sorts of apps are those? I tried testing various models with a test app as a benchmark, a local first app with CRDTs, and many, even frontier models, struggle heavily.

[Kuyawa]

https://mediconsulta.net an AI medical assistant

https://decaboy.fit for tracking progress at they gym

https://megaparley.com sports betting platform

A horse betting platform not published yet, still looking for an API odds provider

A car mechanic AI assistant not published yet

I've learned that the more detailed the initial prompt the better result I get. I can share any prompt if you want

[satvikpendem]

Makes sense, looks like mostly CRUD?

I actually was able to get the CRDT stuff working with the latest models, just had to set thinking to the maximum and also have it continually test itself in a loop via MCP without my intervention so after chugging along for an hour it seemed to have fixed everything itself.

[schaefer]

and you're not alone (I run Qwen 3.6 35B-A3B at home too).

But just for the sake of discussion, let me ask: Who is the service provider you're using to run Deepseek V4? Do you have any way of knowing whether that compute is happening in the US or abroad?