DeepSeek V4 — No Data Retention, Available Everywhere

April 27, 2026

DeepSeek V4 Pro and V4 Flash now run on zero-retention endpoints. The "hosted in China" warning is gone, the China-only routing is gone, and you no longer have to pick them by hand. Auto routing will use them, and Swarm can recruit them, just like every other model on the platform.

What's new

• V4 Pro is now part of the SOTA pool, with a 1M-token context window. Auto mode will pick it for prompts that warrant a frontier model, including very long-context jobs.
• V4 Flash is part of the workhorse pool, also at a 1M-token context window, available as a fast worker on parallel tasks. Especially handy in Swarm.
• Both still support reasoning and tool use, and both are open-weight.
• The model selector no longer shows the "CN" warning dot, and the in-chat data-residency notice no longer appears for either one.

When you'd reach for it

• You're in Auto mode on a long-context job — say, a 400K-token codebase refactor. V4 Pro is now eligible to take it; before today it wasn't.
• You're running a Swarm and want a fast worker for parallel sub-tasks. V4 Flash slots in.

Try it

• "Analyze this 800-page contract and summarize every change request."
• "Audit this codebase and split the work across parallel agents."
• "Give me a structured review of this paper."

Heads up

Pricing is unchanged from the April 24 release — V4 Flash is budget-tier, V4 Pro is workhorse-tier. The earlier note's "ZDR doesn't apply" caveat is now superseded.