DeepSeek
DeepSeek V4: max power, mini price, open weights
🧒 Discovery
The most famous Chinese robot released its new version: very strong, almost free, and its recipe is public.
📰 Essential
DeepSeek releases V4 in two versions (Pro and Flash) with a one-million-token context, thinking modes and open weights anyone can download.
🔬 Expert
The price story is about active parameters: V4-Pro has 1.6 trillion parameters in total but uses only 49 billion per token, V4-Flash 284 billion with 13 billion active. That expert-based architecture is what turns efficiency into a selling point. Two practical notes: the model is labelled "Preview", and the old deepseek-chat and deepseek-reasoner API names are retired after July 24, so developers must migrate.