At the beginning of 2025, DeepSeek, as an emerging force in China's AI field, quickly sparked a global craze with its high - performance, low - cost open - source large - language model (LLM). Its products topped the charts in multiple countries' app stores, becoming a paragon of AI democratization and industry empowerment. This not only highlighted China's innovation capabilities in the generative AI field but also injected new vitality into the global tech competition, propelling various industries into a new era of intelligence.

Against the backdrop of empowering industry digital transformation with DeepSeek's large - model technology, recently, Beijing Jiaxun Feihong Electric Co., Ltd. (hereinafter referred to as “Jiaxun Feihong”) and Huawei Technologies Co., Ltd. (hereinafter referred to as “Huawei”) have jointly created an intelligent railway - industry solution based on DeepSeek, ushering in a new chapter of in - depth cooperation.

The two parties have recently successfully completed the full - stack localization deployment of multiple versions of the trillion - parameter LLM “DeepSeek - R1 - 671B (full - version) / DeepSeek - V3 - 671B (full - version) / DeepSeek - R1 (distilled - version)” based on Huawei's Ascend computing cluster. This achievement not only reflects the collaborative innovation of both parties' technical capabilities but also represents a precise response to the intelligent needs of the railway industry. It marks a new level of in - depth cooperation between Jiaxun Feihong and Huawei in the field of AI large - model technology and injects new momentum into the intelligent development of the railway industry.

In this technical collaboration, Jiaxun Feihong has developed intelligent solutions for four major business scenarios in the railway industry—“production and maintenance, emergency command, inspection and monitoring, and vocational training and education”—based on Huawei's Ascend computing cluster FusionCube A3000 training/inference hyper - converged all - in - one solution.

Additionally, by establishing and efficiently dynamically invoking the DeepSeek knowledge base, Jiaxun Feihong has created an AI - enabled platform for upper - layer business applications in the railway industry. This platform connects the localized deployment environment with the DeepSeek large - model, leveraging DeepSeek's powerful reasoning capabilities to empower upper - layer business applications.

1. Data Enablement

• Data Aggregation and Normalization: Through local data aggregation, the system normalizes multi - dimensional data, including structured, semi - structured, unstructured data, and QA (Question - Answer) pairs.

• Data Preprocessing: The data processing module performs text segmentation, chunk processing, and other preprocessing tasks to generate high - quality QA pairs.

• Knowledge Base Archiving: Through passage analysis and vector - based storage, the knowledge base is archived.

• Knowledge Base Configuration: With a rich selection of knowledge base configurations, the system supports the integration of the DeepSeek large - model and external search engine calls, constructing a three - dimensional knowledge base that includes vector libraries, file libraries, and databases.

2. Application Enablement

• Interactive Modes: Through text input, voice input, and system invocation, the system leverages DeepSeek’s knowledge - base retrieval capabilities to achieve vector - based queries, feature ranking, question construction, conversation sharing, and semantic feature annotation.

• Powerful Inference and Optimization: With the robust reasoning capabilities of the DeepSeek large - model, the system presents multi - dimensional reasoning results and dynamically evaluates and optimizes query results.

3. Model Enablement

• Model Management and Evaluation: The system offers model management and evaluation functions, supporting one - click deployment and management of training and inference services.

• Model Management Capabilities: These include model weight upload, recycling, and quantization format conversion.

• Model Service Capabilities: These include configuration of distributed inference strategies, hyperparameter configuration, and deployment of quantized models.

• Unified Northbound Interface Gateway: The system supports SK authentication and OpenAI - style API calls, and provides API access statistics.

Through the deep integration and full - chain optimization of the above technologies based on the DeepSeek large - model technology, the specific innovative breakthroughs are reflected in the following major application scenarios:

1. Production and Maintenance

The system has enabled several intelligent functions for the railway industry:

• Knowledge Base Construction: By building a railway - industry - specific knowledge base, the system integrates multi - dimensional information such as standards, regulations, equipment documentation, and maintenance data. It deeply analyzes maintenance records, technical documents, and equipment manuals to form a dynamically updated knowledge base, facilitating real - time and efficient queries for maintenance personnel.

• Intelligent Risk Prediction: Based on intelligent analysis of equipment data, the system can accurately identify potential risks, predict trends, and issue timely warnings. In case of equipment anomalies, it quickly locates the root cause of failures by searching historical cases and maintenance documents, improving fault - handling efficiency and reducing maintenance costs.

• Automated Repair Suggestions: Combining historical repair methods and equipment maintenance manuals, the system can automatically generate repair suggestions and provide expert - level guidance, further optimizing maintenance decision - making.

• Interactive Query Capability: With voice or text - based conversational interaction, maintenance personnel can quickly obtain equipment operation parameters and statistical data, completely replacing traditional manual query operations and significantly enhancing maintenance efficiency and decision - making capabilities.

2. Emergency Command

The system has achieved efficient and precise emergency response and decision - making support. Firstly, based on the actual situation of emergency command and in combination with the existing railway emergency resources, social public resources, and emergency response standards and requirements, the system can quickly generate emergency command plans for the scene of emergency response, providing scientific guidance for on - site commanders. Secondly, leveraging DeepSeek’s powerful reasoning capabilities, the system can real - time parse emergency accident scenarios and, in conjunction with knowledge - base data, rapidly derive emergency management and response plans for railway - related accidents. It can also quickly mobilize surrounding emergency resources to minimize traffic accidents and their resulting casualties, property damage, and social impact. This enhances the overall prevention and control level and risk - resistance capabilities of the railway, ensuring the safe operation of transportation and production. Additionally, by associating with the historical accident case library through DeepSeek, matching the latest disposal standards, and using causal reasoning methods to precisely locate the causes of disasters, the system can better understand the root causes of accidents. This provides strong data support and decision - making basis for future prevention and response efforts, comprehensively improving the scientific nature and effectiveness of emergency command.

3. Inspection and Monitoring

The system has achieved comprehensive optimization in lightweight deployment, multimodal fusion, and risk assessment and prediction. Firstly, through DeepSeek distillation and the MoE (Mixture of Experts) dynamic perception system, the system compresses the multimodal large - model knowledge for railway inspection and monitoring scenarios into a lightweight small - model, realizing lightweight deployment and multimodal fusion. Secondly, based on the cross - modal alignment capability of the DeepSeek - VL2 model, the system uses a CLIP - style image - text contrastive learning framework to jointly embed railway slope structure diagrams with text descriptions. It also incorporates a temporal - spatial cross - attention module to align vibration - sensing waveforms (sampled at 50Hz) with video monitoring data (30fps), enhancing the precision of data fusion. Additionally, the system integrates real - time multidimensional perception data, historical accident data, equipment maintenance records, and meteorological environment information to build a multidimensional risk - assessment and early - warning model. It automatically predicts high - risk sections and periods under special weather conditions, allowing for preemptive control measures and emergency plans to be implemented in advance. This significantly enhances the intelligence level of railway inspection and monitoring as well as the capability of risk prevention and control.

4. Vocational Training and Education

The series of technological breakthroughs achieved by this system renewal would not have been possible without a solid technological foundation and comprehensive capability support: Huawei's FusionCube A3000 training/inference hyper - converged all - in - one machine, which is deeply compatible with DeepSeek V3 & R1 and distilled models and supports private deployment, can significantly enhance the efficiency and reliability of AI applications and further accelerate the digital transformation of enterprises.

After testing and selection, the FusionCube A3000 solution, with its core advantages of “light - weight one - stop solution, fast application go - live, strong professional capabilities, and specialized delivery practice,” provides enterprises with full - process support from deployment to application:

The light - weight one - stop solution is reflected in its flexible design, which supports starting with a single node and can be smoothly expanded as needed. It integrates AI computing, general - purpose computing, storage, and networking in a full - stack manner to meet AI application needs in one go.

The fast application go - live is attributed to the pre - adaptation of the DeepSeek large - model, which allows one - click configuration of inference services and significantly reduces deployment time.

The strong professional capabilities are manifested in the built - in full - process toolchain, which, through data engineering tooling, increases the efficiency of knowledge - base construction by 10 times.

Specialized delivery practice is realized through full - stack pre - integration and pre - validation, enabling a quick start - up within 4 hours and ensuring efficient delivery.

Since the signing of the “Comprehensive Cooperation Agreement” in 2023, Jiaxun Feihong and Huawei have fully leveraged their respective strengths, and their cooperation in the field of intelligent transportation has been continuously deepened. Huawei, with its profound accumulation in digitalization and intelligence, especially the concept of building a “point - line - surface - body” digital and intelligent foundation for large - transportation and logistics, has provided solid technological support for the industry. Jiaxun Feihong, relying on its in - depth experience in the railway industry, has accurately grasped customer needs and quickly created customized solutions, demonstrating outstanding capabilities in core scenarios such as safety protection, dispatching command, and operation maintenance. The release of this intelligent solution based on the AI large - model DeepSeek is another important achievement of the deep integration of both parties' technological cooperation and industry experience.

Going alone is fast, but going together is far. The joint creation of Jiaxun Feihong and Huawei is not only a deep integration of railway - industry scenarios and AI large - model technology but also a strategic choice for both parties, based on their profound insights into the digital and intelligent transformation of the industry, to jointly promote the intelligent development of railways. By reshaping the intelligent boundaries of railways with the DeepSeek large - model and penetrating the underlying hardware - technology barriers with Huawei's Ascend, the jointly - created solution will fully leverage the multiplicative effect of each party's advantageous resources. Together, they will build an end - to - end capability for railway - industry customers, representing a deep integration of the railway industry and AI large - model technology and injecting strong momentum into the digital and intelligent transformation of the railway industry.