Huibo Investment Research recently released a research report to comment on the artificial intelligence company DeepSeek, the main contents of which include: DeepSeek's development status, market performance, technological innovation, application scenarios, industrial opportunities, related company analysis and future prospects.

DeepSeek was founded in 2023 and is a subsidiary of High-Flyer Quant. Since its establishment, it has continuously developed and iterated large models, and its product system has become increasingly rich.

For example, DeepSeek-v3 tops the list of open source models in the mainstream list of large models, and is on par with the world's advanced closed-source models. The performance of the inference model R1 is benchmarked against the official version of OpenAI's o1, and it performs well in many fields. Moreover, DeepSeek focuses on "extreme cost performance", with a training cost of only $5.576 million, which is one-tenth of GPT-4o, and the cost of API calls is only one-thirtieth of OpenAI. At the same time, technology giants such as NVIDIA, Microsoft, and HUAWEI CLOUD have connected to DeepSeek to promote the development of the global AI ecosystem.

In terms of market performance, DeepSeek-V3 has surpassed many open-source models in a number of evaluations, and has performed well in knowledge tasks, long text evaluation, algorithm code scenarios, etc., with the generation speed of words increased to 60TPS, and the pricing of API services is also very cost-effective. The R1 model is comparable to OpenAI o1 in terms of inference capabilities, open-sourced multiple models and supports "model distillation", launches APIs and apps, and the service pricing is much lower than OpenAI o1. After the launch of the two models, DeepSeek became the world's fastest-growing AI application, with 15 million daily active users in the 18 days after it was launched, a growth rate 13 times that of ChatGPT.

Technological innovation is the core competitiveness of DeepSeek. It uses model distillation technology to improve the inference ability of small models, which is better than reinforcement learning, and also unifies multimodal understanding and generation through visual decoupling. In terms of architecture design, innovative architectures such as Multi-Head Latent Attention (MLA) and Deep Seek Hybrid Expert System (DeepSeekMoE) reduce memory footprint and computational load.

At the same time, multi-level technical improvements are carried out at the model layer, architecture layer, training layer and inference layer, such as the use of MoE architecture, the introduction of load balancing strategy without auxiliary loss, the use of DualPipe algorithm and FP8 mixed-precision training, and the migration of R1 inference ability, etc., so that the performance and efficiency of the model have been significantly improved.

DeepSeek's technological innovations have driven the adoption of AI across multiple industries.

In the field of AI + film and television, it can reduce the cost and time of film and television production, and has applications from script generation to post-production; In terms of AI+ games, it can improve the efficiency and experience of game development, automatically generate game assets, and optimize the rendering effect; In AI + social companionship, virtual assistants and characters can provide emotional support and personalized services; AI+ e-commerce can achieve accurate recommendation and automated customer service to optimize operations; In the field of AI + marketing, personalized marketing strategies can be realized and high-quality marketing content can be generated.

In terms of industry opportunities, the open-source nature of DeepSeek has accelerated the adaptation of AI industry chain enterprises, and cloud vendors have launched their models one after another. The "Jevons paradox" shows that its technological breakthroughs may lead to an increase in computing power demand, and third-party cloud factories are expected to benefit from model equality. At the same time, DeepSeek promotes the cost compression of end-side inference, and has broad application prospects in end-side devices such as AI glasses, headsets, learning machines, and toys.

Among the related companies, domestic computing power and computing power service companies such as Runjian Co., Ltd. (002929), Sugon (603019), Haiguang Information (688041) and other active layout. Runjian Co., Ltd. deployed DeepSeek to empower intelligent twins application development, and the computing power business developed well; Sugon's performance is stable, and the computing industry ecology continues to improve; Haiguang Information DCU products were rapidly iterated and the DeepSeek model adaptation was completed.

AI application companies such as Kingsoft Office (688111), Color News (300634), and Straight Flush (300033) are also developing with the help of DeepSeek. Coordinated development of intelligent, localized and cloud-based Jinshan office; Color News Co., Ltd. lays out Rich AI super factory; Flush is deeply engaged in financial information services, and lays out AI large models and applications.

Looking to the future, DeepSeek is expected to promote the arrival of the era of AI equality, accelerate the development of applications, and its open-source large model ecosystem may become Android in the AI era. It will also promote the prosperity of the AI ecosystem, achieve high-quality model parity, raise the lower limit of model capabilities, and accelerate the iteration of the AI industry. However, as the United States continues to increase AI export controls and restrict the use of DeepSeek, China's AI industry is facing chip bottlenecks, and it is crucial to break through advanced manufacturing processes and achieve independent and controllable semiconductors.