On April 24, DeepSeek, a leading force in Chinese AI, officially released and open-sourced its fourth-generation models: DeepSeek-V4-Pro and DeepSeek-V4-Flash. This series represents a significant leap in technical architecture, expanding the context window from 128K to 1M.
The models lead the open-source field in Agent capabilities, logical reasoning, and world knowledge. The Pro version features a 1.6T parameter scale, targeting high-end scientific research and complex production scenarios, while the Flash version, with 284B parameters, focuses on cost-effectiveness and rapid response, matching the performance of top-tier overseas models in its class.
Simultaneously, Huawei’s Ascend Computing Business Department announced that through deep "chip-model synergy" technical breakthroughs with DeepSeek, the full range of Ascend Supernode products now fully supports DeepSeek-V4. As the core of domestic computing power, the Ascend 950 chip effectively reduces the Attention calculation and memory access overhead associated with ultra-long contexts by integrating Kernel and multi-stream parallel technologies. Combined with multi-dimensional quantization algorithms, it achieves high throughput and low latency in large model deployment. Additionally, the Ascend A3 Supernode, designed for ultra-large-scale training, has completed adaptation and provides a complete training reference implementation, significantly lowering the barrier for enterprises to fine-tune V4 models using domestic hardware.
Industry analysts believe that the deep integration of DeepSeek-V4 with Ascend computing power marks the end of the Chinese AI industry's reliance on overseas chips, officially entering the era of "native hardware adaptation." Boosted by this news, domestic chip concept stocks and the Huawei computing power supply chain saw a collective surge in afternoon trading, with multiple leading stocks hitting their daily price limits. Although DeepSeek stated that online service throughput for the V4-Pro version remains limited due to high-end hardware constraints, the supply-demand imbalance for domestic computing power is expected to ease substantially in the second half of the year as Ascend A3 and 950 clusters are deployed at scale.