DeepSeek-V4

DeepSeek's fourth generation keeps frontier weights under plain MIT while splitting the family into two lines: V4-Pro, a 1.7 trillion parameter mixture-of-experts flagship whose 0813 build left preview in mid-August 2026, and the lighter V4-Flash aimed at cheaper high-volume serving, both preceded by preview releases. The models expose three reasoning effort levels, low, high, and max, with recommended output budgets reaching 384K tokens at the top setting, and they ship with DSpark, a speculative decoding module that accelerates generation during agentic coding and multi-turn reasoning sessions. Serving guidance targets vLLM with DSpark enabled or SGLang with the FlashInfer backend, using FP8 KV caches to hold memory down; the Pro model is sized for a four-way GB300 node with expert-parallel and data-parallel layouts. For teams that cannot host it, DeepSeek's own API mirrors the open checkpoints, but the MIT license means clouds and enterprises can serve, fine-tune, and resell the weights without restriction.