DeepSeek V4 tests fast, expert, vision modes

If this report is accurate, DeepSeek is signaling a more productized model stack: users get explicit modes instead of guessing which model variant to call. For developers, that changes prompt design, routing, and evaluation because mode choice now becomes part of the interface contract.

• –Fast mode looks like the default path for cheap, low-latency work
• –Expert mode suggests deeper reasoning and better agent-style workflows
• –Vision mode implies the multimodal layer is being exposed as a first-class option
• –A gray-scale rollout usually means DeepSeek is still validating load, quality, and safety before a broader launch
• –The Huawei-chip reporting matters because it hints the release is as much about inference infrastructure as model quality