DeepSeek is back with V4, slashing agentic AI costs

DeepSeek introduced V4, an open-weight model built for agentic workflows and long-context tasks.

The company says V4 uses far less compute and KV cache than its predecessor while still reaching top benchmark results.

V4 is optimized for Huawei Ascend chips, which could lower deployment costs and reduce reliance on U.S. hardware.

If the performance and efficiency claims hold, DeepSeek could pressure proprietary AI leaders and strengthen non-U.S. AI infrastructure.