Bluesky Thread

ok, so DeepSeek V3.1 does appear to be real. 128K context. unsure what other ...

August 19, 2025 View original thread

ok, so DeepSeek V3.1 does appear to be real. 128K context. unsure what other details are real since they don’t announce the same way western labs do. the name might actually be a date checkpoint too, idk

it’s in the web version, appears to be a reasoning model, and the reviews are NOT GOOD, even from deepseek sycophants

Teortaxes
@teortaxesTex
（DeepSeek 推特铁粉202...
I don't know what is happening. This model seems 1) not clearly smarter at a glance 2) inclined to think for MUCH shorter (40-80% cut) and 3) starting 100% of CoTs with "Hmm" (usually followed with", the user"). I'd say we're hallucinating but CoT reduction is consistent.

More like this