Nobody wants huge "Pro" models.

I think, moving forward, DeepSeek will only produce flash tier models. Internally DeepSeek V4 Pro underperformed really hard and nobody was using it, while DeepSeek V4 Flash got so insanely popular that they could barely keep the servers running. So what's the point of keeping creating and deploying huge 1 trillion parameter models if demand is not there? My hypothesis is simple. The people have spoken. They want cheap, reliable beasts that you can plug into some kind of harness extension CLI or whatever, and it will reliably call tools and perform some basic operations. My guess is we won't see anything above 800b parameters anytime soon. Source: I saw it in my hallucinations while I was blackout drunk.