China Academy of Information and Communications Technology Officially Launches DeepSeek V4 Localization Adaptation Testing
The China Academy of Information and Communications Technology (CAICT) announced today the official launch of DeepSeek V4 localization adaptation testing, promoting in-depth collaboration between the model and domestic software and hardware, and accelerating industrial implementation. The test is based on the key laboratory of the Ministry of Industry and Information Technology and the AISHPerf benchmark system, covering a full stack of AI software and hardware products including chips, servers, integrated machines, clusters, development toolchains, and intelligent computing platforms, focusing on the inference and fine-tuning processes of the entire DeepSeek V4 series of models.
The evaluation assesses from five dimensions: adaptation ease of use, functional completeness, optimization effect, performance, and cost. It also adds specialized indicators such as long sequence processing, code capability, agent call success rate, and task decomposition, forming a three-dimensional evaluation system.
Multiple domestic hardware companies achieved Day-0 adaptation on the day of DeepSeek V4’s release, marking the entry of domestic AI software and hardware into a stage of synchronous iteration.
This test will objectively verify the adaptation level, strengthen domestic computing power support, and accelerate the construction of a self-controllable AI ecosystem.
DeepSeek V4 includes two versions: V4-Pro (flagship version) and V4-Flash (lightweight version), both natively supporting ultra-long contexts of 1 million Tokens (approximately 750,000 characters). They adopt a self-developed DSA sparse attention mechanism, reducing inference costs for million-context by 70% and reducing memory usage by 40%.
V4-Pro: With a total of 1.6 trillion parameters and 49B activated parameters, it focuses on top-level performance, benchmarking against globally leading closed-source models such as GPT-5 and Claude Opus, and is suitable for complex inference, code generation, and scientific research computing tasks.
V4-Flash: With a total of 284B parameters and 13B activated parameters, it focuses on efficiency and low cost. Its inference capability is close to the Pro version, but it is faster and cheaper, suitable for daily interaction, content creation, and lightweight enterprise deployment scenarios.