On: DeepSeek-v3.2: Pushing the frontier of open large language models [pdf]

[cherioo]

Allegedly deepseek is doing this because they don’t have enough gpu to serve two models concurrently.