TRS has always adhered to the two-wheel drive strategy of "high-quality data + industry large model", and firmly believes that open source will become the mainstream of basic large model. Since the release of TRS-GPT LLM in June 2023, in terms of data service, TRS's high-quality data set has been adopted by dozens of companies and research units such as Xiaomi, Huawei, Yunzhisheng and Shanghai Artificial Intelligence Laboratory; In terms of application of large models, seven major industry models of government affairs, media, finance, public opinion, public security, open source intelligence and patents have been launched successively, and the number of application cases ranks the leading position in China; in order to reduce the deployment and delivery cost of large models and increase flexibility, the TRS-GPT LLM Integrated Platform has been launched.

At the beginning of 2025, DeepSeek trained a large open source reasoning model comparable to OpenAI o1 with its high performance and low cost advantages, which attracted worldwide attention. For the integration with DeepSeek, TRS does not stay at the primary level of adaptation and use, but with the triple advantage of "platform + system + service", fully integrates DeepSeek, comprehensively improves the customer's large model use experience and reduces the cost of large model projects, greatly enhancing the core competitiveness of TRS-GPT LLM.

Platform empowerment: building a more powerful TRS-GPT LLM

With reference to DeepSeek's technical route, TRS further optimized and enhanced its self-developed "TRS-GPT LLM". It integrates 30 years of technical accumulation in NLP field, more than 10 years of high-quality data and knowledge asset construction experience, and the application practice of more than 10000 enterprise users. By introducing DeepSeek's advanced technologies, such as Hybrid Expert (MoE) architecture and industry knowledge enhancement technologies, TRS-GPT LLM become more computationally efficient and flexible. The new version of TRS-GPT LLM adopts the hybrid architecture of "universal base + vertical experts", realizes parameter-level task allocation through dynamic routing algorithm, and adapts the results of multiple open source large models to migrate to the industry service field through DeepSeek training optimization.

System upgrade: TRS industry software full access to DeepSeek

TRS has mature industry solutions in many fields such as digital government, financial media, financial technology, public security, etc., fully accesses DeepSeek technology, creates DeepSeek-R1 + mass industry data + industry application solution, and realizes deep integration and upgrading at system level. Industry-oriented enterprise-level AI services focus on creating explosive system applications by adapting to the authenticity needs of end users, coupling high-quality data and industry cognition in the mode of Agent Agent agents, and with long-term cooperative customers. Through this deep integration, TRS's industry solutions have stronger intelligent capabilities, can handle complex business scenarios more efficiently, create greater value for customers, and achieve implementation at lower cost.

Service optimization: SaaS services fully support DeepSeek

The five major data asset platforms of TRS: Media Information, Public Opinion, Financial Technology, Open Source Intelligenceand Global Insight have accumulated more than 300 billion big data resources with high commercial value. On this basis, TRS fully introduced DeepSeek technology to further optimize SaaS services. It not only reduces the application cost of large models through MLA and MoE architecture, but also realizes accurate data indexing and data value mining; It also embeds automatic scheduling agent agent in the data SaaS service provided for customers. Users can independently select DeepSeek, Llama, Qwen and other large models to construct dual engines combining high-quality data and large model capability, which not only further increases the controllable generation of Chinese, but also provides customers with more personalized and efficient data services.