DeepSeek Unveils V4 Grey Scale Test and Expert Mode

On April 7, Readhub reported that DeepSeek V4 is undergoing intensive gray-scale testing. Screenshots shared by multiple programmers and tech bloggers reveal that the new model generation has achieved breakthroughs in its underlying architecture, alongside significant upgrades to its interaction logic and multimodal capabilities.

Key Highlights: A Three-Tiered New Functional Architecture

Based on the leaked test interface, DeepSeek V4 offers three core options, signaling a comprehensive evolution of its product lineup:

Lite Version (DeepSeek V4 Lite): Prioritizes response speed, ideal for everyday, lightweight conversations.

Expert Version (DeepSeek V4): A mode focused on deep logical reasoning, potentially built upon the "new memory architecture" referenced in a paper co-authored by Liang Wenfeng.

Vision Version (DeepSeek V4 Vision): Represents the deep integration of multimodal capabilities, enabling direct handling of image and video analysis tasks.

Technical Approach: A Firm Advocate for "Domestic Chips"

The rise of DeepSeek is driven not only by algorithms but also by deep integration with the local computing ecosystem:

Prioritizing Domestic AI Chips: Reports indicate DeepSeek is developing at least two large models fully based on domestic AI chips, having already initiated a domestic chip procurement surge.

Refusing External Dependence: In previous testing applications to US chip manufacturers, DeepSeek did not grant access to V4 model testing, instead prioritizing domestic companies for collaborative optimization.

Industry Expectations: Can It Challenge the Pinnacle of Programming?

Beyond general performance gains, the market holds specific vertical expectations for DeepSeek V4:

AI Programming Special Edition: Industry speculation points to a version tailored for code generation and engineering implementation, positioning it to compete directly with Anthropic's Mythos or OpenAI's GPT series.

Ultra-Long Context Processing: Building on DeepSeek's established strengths, V4 is anticipated to handle ultra-long contexts of millions of tokens in a single pass.

Topic Tracking: A Long-Awaited Release Cycle

From the leak of the MODEL1 new architecture in January, to the February release of the OCR 2 model, and now frequent gray-scale tests, the launch of DeepSeek V4 appears imminent. As the expected release window approaches this month, this flagship model—emphasizing "native memory" and "domestic adaptation"—may once again redefine the price-to-performance benchmark for domestic large models.

Conclusion: An AI That Better Understands Chinese Computing Power

From automated code validation to novel visual interactions, DeepSeek is demonstrating that domestic models can achieve remarkable evolutionary progress without relying on top-tier overseas chips.