The China Academy of Information and Communications Technology announced today that it has officially launched the localization adaptation test of DeepSeek V4 to promote in-depth collaboration between the model and domestic software and hardware and accelerate industrial implementation.This test was carried out based on the Key Laboratory of the Ministry of Industry and Information Technology and the AISHPerf benchmark system, covering full-stack AI software and hardware products such as chips, servers, all-in-one machines, clusters, development tool chains, and intelligent computing platforms, focusing on the inference and fine-tuning processes of the full series of DeepSeek V4 models.
The evaluation evaluates from five dimensions: ease of adaptation, functional completeness, optimization effect, performance, and cost, and newly adds special indicators such as sequence processing, code capabilities, agent call success rate, and task disassembly to form a three-dimensional evaluation system.
On the day of its release, DeepSeek V4 has achieved Day-0 adaptation for multiple domestic hardware, marking the entry of domestic AI software and hardware into the same-frequency iteration stage.
This test will objectively verify the adaptation level, strengthen the support of domestic computing power, and accelerate the construction of an independent and controllable AI ecosystem.
DeepSeek V4 includes two versions: V4-Pro (Flagship Edition) and V4-Flash (Lightweight Edition). Both versions natively support 1 million Token ultra-long contexts (approximately 750,000 words). They adopt a self-developed DSA sparse attention mechanism to reduce the cost of inference for one million contexts by 70% and reduce memory usage by 40%.
V4-Pro: The total parameters reach 1.6 trillion, and the activation parameters are 49B. It focuses on the top performance limit, benchmarks against the world's top closed-source models such as GPT-5 and Claude Opus, and is suitable for complex reasoning, code generation, scientific research calculations and other difficult tasks.
V4-Flash: Total parameters 284B, activation parameters 13B, focusing on high efficiency and low cost, reasoning ability close to the Pro version, faster speed, lower price, suitable for daily interaction, content creation, enterprise lightweight deployment and other scenarios.
