The technical team of Sinopec worked overnight to tackle the problem, relying on their self built domestic intelligent computing cluster. On the first day of release, they completed the privatization deployment of the model and connected it to the Great Wall Artificial Intelligence Platform to provide online services for the entire group.

A zero day response to a race against time

On April 24, 2026, while the domestic AI community was still discussing the news of the release of the DeepSeeker V4 preview version, engineers from the Information Department of Sinopec had already entered a practical state. For this technical team that has been deeply cultivating the domestic intelligent computing resource pool for many years, this is not only a version update, but also a "practical exercise" to test their independent and controllable capabilities.

We have been closely monitoring its compatibility with domestic chips since the model architecture design stage, "said the technical director of Sinopec. In the first time we obtained the weight of the open-source model, the team relied on its self built domestic intelligent computing resource pool to carry out intensive deployment and optimization work. The entire process from model loading, inference testing to performance stress testing is completed in a closed loop within a few hours. On the day of its release, the DeepSeek-V4 Flash model successfully integrated with Sinopec's unified Great Wall Artificial Intelligence platform, achieving a "zero day response" from open source release to enterprise privatization online. Currently, the model has provided inference services to the entire group.

Hard core breakthrough of technological foundation

The DeepSeeker V4 Flash model chosen by Sinopec this time, although an economical version, has no reduction in its technological value. This model adopts a hybrid expert (MoE) architecture with a total parameter count of 284 billion and native support for an ultra long context window of 1 million tokens. It has stronger coherence when dealing with complex petroleum geological documents or long chain code generation tasks.

More importantly, it deeply adapts to the characteristics of domestic computing power. DeepSeeker V4 synchronously validated the Huawei Ascend NPU platform during the design phase, and solved the performance bottleneck of large models running on domestic chips through targeted operator optimization and inference framework adaptation. For state-owned enterprises like Sinopec that emphasize supply chain security and independent controllability, this "chip model" collaborative localization ecosystem is precisely the confidence they have to deploy on a large scale on the first day of release.

AI driven secondary entrepreneurship in industries

This lightning deployment is not Sinopec's first AI trial. In recent years, Sinopec has deeply implemented the national artificial intelligence+strategy and has built a complete digital intelligence system from bottom level computing power to upper level applications. The previously independently developed and officially certified Great Wall model (including 700 billion and 70 billion parameter versions) has accumulated rich industrial knowledge in scenarios such as oil and gas exploration and refining process optimization.

The entry of DeepSeeker V4 will inject stronger general intelligence into this system. It is not only a "super assistant" for employees' daily office work, but also a "strong brain" for the upcoming release of the FiberHome industrial intelligent agent. In the future, in core businesses such as drilling design, equipment failure warning, and new material research and development, this domestically produced AI base will undertake more complex logical reasoning and code generation tasks, promoting Sinopec's transformation from a traditional energy company to a technology company that combines artificial intelligence and energy and chemical engineering, injecting new momentum into the secondary entrepreneurship of the entire industry chain.Editor/Yang Meiling