DeepSeek-V4
The Chinese laboratory released a family with 1.6 trillion parameters in the larger model and a million-token context window, under an MIT licence.
Why it matters
Open weights stopped trailing even by months: the difference from frontier closed models narrowed to the price of a request.
Pro holds 1.6 trillion parameters with 49 billion active per token, Flash 284 billion with 13 active. Both carry a million tokens of context. On agentic measures the laboratory places the larger model alongside frontier closed ones, which is its own claim rather than an outside measurement. The MIT licence carries no restriction by service size or purpose. This is the third year running in which a Chinese laboratory sets what open weights means.