DeepSeek V4

Editor's Review

DeepSeek V4 is real, shipping, and the most aggressive value play in frontier AI — a 1.6T-parameter MoE (49B active per token) with 1M-token context, 384K max output, and MIT-licensed open weights on Hugging Face. The family: V4-Flash-0731 (July 31, 2026, 284B MoE), V4-Pro-0813 (GA August 13, 1.6T), and V4.1-Flash (September 10, 552B encoder-decoder MoE with native image input, which retired V4-Flash the same day). API pricing introduced a peak/off-peak split on August 17, 2026 — Flash at $0.44/M input tokens at peak, exactly half off-peak, cache hits near zero. The consumer chat app is free. The differentiator is the open-weights story at this scale: Artificial Analysis scored the 0813 Pro build 53, ranking #3 on its leaderboard, from a model you can self-host under MIT. The sparse-activation engineering (8B active on prefill for the 1.1-Flash) is how it prices the cheap tier so aggressively. The caveats are real. Independent benchmarks landed a few points below DeepSeek's own model-card claims, and the API lineup churned twice in two months (V4-Flash retired, V4-Pro nearly rerouted) — pin model IDs in config, not code, because you don't control the roadmap. Versus Gemini Flash (Google's TPU-infra cheap tier) and Claude Haiku 5.5 ($1/$5 per MTok, launched early October 2026): three cheap-tier philosophies — DeepSeek (sparse MoE), Google (infrastructure), Anthropic (tiered). For self-hosters and cost-sensitive API users, V4 Pro is the pick; for stability, the Americans.