DeepSeek-V4 Pro/Flash ship as MoE models with 1M context and hybrid attention for cheaper agentic inference

LLM-Stats·medium signal

DeepSeek released DeepSeek-V4-Pro and DeepSeek-V4-Flash, MIT-licensed Mixture-of-Experts models with a 1M-token context window and a hybrid attention mechanism aimed at cutting inference cost for long-running agentic tasks (Pro ~1.6T total / 49B active; Flash ~284B / 13B active). With the highest reported open-weights SWE-bench Verified score (~80.6%) and Flash output around $0.28/M tokens, it is a strong cost/capability option for self-hosted agent stacks.