Skip to content
X · @op7418

DeepSeek V4 detailed official announcement is out

DeepSeek V4详细的官方公告也发了:

DeepSeek says V4 Pro has 1.6T total parameters with 49B active, while Flash has 284B total and 13B active; both were pretrained on 32T tokens. Web and app Expert mode map to Pro, and Fast mode maps to Flash. The post also says several benchmarks are on par with Opus 4.6, with stronger agent ability and world knowledge, plus a new attention mechanism that reduces compute and memory demand.

Why it matters: This is a flagship DeepSeek release, scored on par with peer US lab model launches. HKR-H/K/R all pass on concrete scale numbers, 32T data, and an inference-efficiency mechanism; benchmark setup, pricing, and API availability are not disclosed in the summary.

Read the original ↗Export Markdown