DeepSeek V4-Flash

DeepSeek-V4-Flash-0731, officially released July 31, 2026, is the efficiency tier of the V4 family: 284 billion total parameters with roughly 13 billion active per token, a one-million-token context window, and 384K maximum output. The Hugging Face listing shows 304B because it counts the separate DSpark speculative-decoding module alongside the 284B backbone.

Both code and weights ship under the MIT license — genuinely open, permitting commercial use and redistribution. API pricing runs $0.15/$0.60 per million input/output tokens off-peak (exactly half the peak rates), with cache hits priced near zero, making it one of the cheapest ways to run million-token-context agents. It was superseded for flagship workloads by V4.1 Flash in September 2026, but remains a strong open-weight efficiency pick.