After a late-night update, DeepSeek reportedly said: I am V4?

DeepSeek深夜更新后自曝：我是V4（？！）

DeepSeek added Fast and Expert modes on its web app and started gray-testing a Vision model; the claim that Expert mode is V4 comes only from user probes and the model’s own replies. The post gives one concrete detail: Expert mode focuses on code, web, and harder generation tasks, is supply-limited, does not support multimodal or file upload, and one user reported a length cap at about 133K tokens. What matters is the official model ID and context spec; the post does not disclose them, pricing, or a release timeline.

Why it matters: HKR-H is strong on the 'I am V4' hook. HKR-K and HKR-R pass because the post gives testable mode behavior and a ~133K token limit, and DeepSeek silent swaps are highly discussable. The score stays in the mid-70s because the model name, price, and context window remain unconfirmed