DeepSeek V4 Gray Test Launched: New Fast and Expert Modes, Response Speed Significantly Improved

DeepSeek V4 has finally entered the gray test phase. Users on the web and some APP versions noticed two new modes: Fast and Expert. The Fast mode is suitable for daily conversations, with instant responses and support for image/file recognition; the Expert mode focuses on complex problems, performing better in tasks like physical simulation and mathematical reasoning, but requires waiting during peak hours.

Both modes are equipped with a knowledge base up to May 2025 and a 1M token context window. User tests show a significant improvement in response speed. In legal text processing, it not only has strong retrieval capabilities but also provides specific operation suggestions; it also gives detailed guidance on Excel-related issues.

The new version hasn't been fully rolled out yet. If users don't see the update, they can enable it by entering "Expert Mode" in the chat box. However, some users found that the Expert mode performed worse than Qianwen on certain physical problems (like a long bamboo pole passing through a city gate), and there are speculations that it might still use the V3.2tk version; some gray test users also saw a "Visual Mode" option, but its specific architecture is not yet clear.