DeepSeek V4 Flash (Latest)

DeepSeek-V4-Flash-Latest is not a separate model but a rolling alias: providers re-target it to the newest DeepSeek-V4-Flash checkpoint whenever one ships, so calling this id always gets the current release without any code change on your side. Pinned checkpoints like the original release or the 0731 build stay immutable next to it. The trade-off is the standard one across AI labs: use the alias in development or wherever you always want the newest behaviour, pin a dated checkpoint in production where reproducibility matters, because the model behind this id changes under you by design. Capabilities track whatever checkpoint the alias currently serves, which for the V4 Flash family means a lightweight 284-billion-parameter sparse MoE with 13 billion active parameters, a one-million-token context window, selectable reasoning modes, and tool calling.