Something just evolved on Deepseek

Posted by Omnimum@reddit | LocalLLaMA | View on Reddit | 9 comments

[-]

Reasonable-Rich-5373@reddit

In Instant+DeepThink mode, the generation has become much better.
I have long been testing with this prompt:

Build a fully holographic badass html single code web page about DeepSeek 3.1 go all out and add the things I didnt ask for . More holograms. Make sure it is functional and unique . Do not worry about size restraints as we can continue from where ever you may stop

[-]

paperbenni@reddit

My guess is this is DS 3.2 with different reasoning budgets, meaning we'll have to pay for the full version soon

[-]

Artistedo@reddit

It is another model. The thinking process is significantly, and I mean a lot different to the previous one. I would be surprised if this isnt a complete train from scratch

[-]

Equivalent-Word-7691@reddit

I hope this isn't v4 otherwise is a HUGE let down 😅

[-]

nullmove@reddit

If the expert is V4 proper, we are utterly cooked. Though instant does feel great if this is indeed some ~280B lite model.

[-]

This is not DS4, model still feels same and fails the same tests, frontier models can do. Also reports same knowledge cutoff and no multimodality that DS4 will have... but looks like they are preparing for a launch of some kind.

[-]

BagelRedditAccountII@reddit

Reportedly, this is a soft-launch of Deepseek V4. Glad to see some tangible evidence of an update that isn't pure hearsay.