a devlog on machines & languages

2 DeepSeek V4 Flash

This month, I started working on building Mogra, an OCR engine to handle Devanagari and Latin that does not involve neural networks so that it can run fast even on a toaster. I had some credits lying around on the DeepSeek platform and decided to make use of it. I haven't really used the API all that much since my project in early 2025 when I translated about 60K pages of public domain English classics into Sanskrit. And DeepSeek V3 hasn't really impressed me as a coding assistant.

But... I was pleasantly surprised when DeepSeek V4 Flash proved me wrong. Not only was it smart, but it was blazing fast, and the pricing was incredibly attractive:

TypePrice
Cached Read$0.0028/M
Input$0.14/M
Output$0.28/M

Until DeepSeek bumped prices up yesterday, that is.

TypeOff-peakPeak
Cached Read$0.007/M$0.014/M
Input$0.22/M$0.44/M
Output$0.66/M$1.32/M

I got a lot of work done with it and renewed my OpenCode Go subscription when the DeepSeek credits ran out.