DeepSeek previews V4: 1.6T params, 1M context, 75% API price cut

DeepSeek-V4 preview ships with 1.6T total / 49B active MoE parameters, a 1M-token context window, and a 10x smaller KV cache tuned for agentic workloads. A V4-Flash variant offers 284B/13B. API pricing dropped 75% and cache costs 90%; benchmarks don't meaningfully close the frontier gap.