Million-token context

The V4 preview’s structural claim is the one I use: 1M context is the default on official DeepSeek services, with sparse attention aimed at making that cheap enough to ship. I stopped chopping a 180-page RFC into three chats. The model still hallucinates closures; I just spend less time restating the appendix.

Max output on the API table was 384K. That is not infinite. I still ask for lists and owners, not a rewrite of the whole packet in one blob.

Models: Flash and Pro. Related group: API Platform.