DeepSeek Unveils Open-Source Flagship Model DeepSeek-V4 with In-Depth Adaptation to Domestic Computing Power

On April 24, DeepSeek unexpectedly launched its new generation flagship large model DeepSeek-V4 and made it open-source. The new model matches the world’s top closed-source models in terms of inference performance, while continuing its cost-effectiveness advantage, attracting widespread attention with extremely low inference costs and standard high-performance long-text capabilities.

Notably, DeepSeek’s new launch places greater emphasis on in-depth adaptation to domestic computing power platforms. Analysts point out that the model has prioritized adaptation to domestic chips such as Huawei Ascend, pushing the collaboration between domestic large models and domestic computing power bases to a new high. This move is particularly significant against the backdrop of the critical period for the construction of the domestic chip ecosystem. Securities Daily reported that SoftStone Power’s subsidiary SoftStone Huafang, based on the “Kunpeng + Ascend” dual-engine architecture, took the lead in completing full-stack adaptation and verification of DeepSeek-V4 on the same day of its release.

Coincidentally, Nvidia CEO Jensen Huang mentioned in an interview some time ago that if DeepSeek’s new model made its debut and achieved the best performance on China’s AI hardware platforms, the model’s proliferation would drive Chinese technology to become a global standard. Now, DeepSeek-V4 has indeed made the active choice to adapt to domestic chips.

This choice is not an easy one. Compared with the mature international ecosystem, domestic chips still need continuous improvement in tool chains, operator libraries and system collaboration. Full-link optimization from training to inference often means a longer cycle and higher investment, as well as a willingness to take a backseat in terms of visibility for a period of time. This trade-off reflects a rare innovative mindset.

In the past two years, the large model industry has seen frequent new model launches and financing news, shrouded in an atmosphere of urgency where “no news means being eliminated from the game”. In such an environment, DeepSeek’s choice is slightly different: over the past 15 months, apart from a few technical articles, the team has rarely made public statements, focusing their energy on technical polishing. It was not until the release of this new model that they released phased achievements in a concentrated manner.

In this launch, the team quoted Xunzi’s words: “Not tempted by praise, not afraid of criticism, follow the path and uphold integrity”, which can be regarded as a response to external praise and criticism. Neither glory nor criticism has affected its focus on the main technical direction and its commitment to moving forward towards the established goal.

The launch of DeepSeek-V4 is not only a leap in model capabilities, but also a vivid example of China’s technological innovation. Its focus on in-depth adaptation to domestic computing power and careful polishing of key technologies reflects an innovative attitude of maintaining strategic determinism in a complex environment and continuously exerting efforts in key areas. It conveys a consciousness of not being trapped by short-term evaluations, encouraging more people to “follow the path” and continuously write the “Chinese answer sheet” of independent innovation.