calotterypowerball| Magic Square Quantification DeepSeek-V2: 236 billion parameter AI model, performance exceeds GPT-4, and the API price is only 1% of it.

Newsletter summary

Release of the new DeepSeek-V2 modelCalotterypowerballParameter 236 billion, the performance is comparable to that of GPT-4, and the computing cost is greatly reduced. The size of the AI search market is expected to reach 329.Calotterypowerball.3.5 billion, with a compound annual growth rate of 32.93%. Pay attention to the investment opportunities of related companies.

Text of news flash

Magic square quantization AI subsidiary released the high-performance large model DeepSeek-V2, the market and investors can pay attention to the potential impact of magic square quantization's AI company depth search (DeepSeek) recently launched its second generation MoE big model DeepSeek-V2. The number of parameters of the new product reached 236 billion, and the comprehensive Chinese language ability surpassed GPT-4 in the AlignBench benchmark test. At the same time, the performance of this model is similar to that of LLaMA3-70B in MT-Bench English comprehensive ability test, and is better than Mixtral8x22B, the open source model of MoE. This development is likely to have a significant impact on investors in the secondary market and deserves close attention.

DeepSeek-V2 helps to optimize the computing cost of the large model. The release of the innovative database and algorithm DeepSeek-V2 is based on the high-quality, multi-source pre-training corpus of 8.1T token, which significantly increases the proportion and quality of Chinese data. The MLA architecture innovation of the model greatly reduces the amount of computation and reasoning memory, and the self-developed Sparse structure DeepSeekMoE further reduces the computing cost. In terms of price, DeepSeek-V2 's API pricing is very competitive, with an input cost of 1 yuan and an output of 2 yuan per million tokens, which is only nearly 1% of the GPT-4-Turbo price.

The diversification of AI model market promotes the development of ecological construction MoE model (mixing of experts), which establishes a diversified path for AI model. The MoE model consists of multiple submodels, each of which focuses on a specific subset of the input space. This architecture reduces the interference between sample types through gated networks. At the same time, although the Transformer model performs well in dealing with sequence-to-sequence tasks, its training process requires great computing power and time. Therefore, the new sustainable model with MoE architecture is gradually becoming an industry trend, bringing new growth momentum to the market.

There is a new growth trend in the field of domestic AI products. The growth rate of some segments of domestic AI products is significant, especially in the field of AI search. 360AI search and Secret AI ranked top5 in the domestic acceleration list, with 360AI search growing by 1303.09% in April. It is reported that 360AI search takes the AI search function as the core, provides reference materials, thought guidance and extended reading, and supports PDF documents, audio and video analysis and web page AI collation. According to Leopard's forecast, by 2027, the domestic AI search engine market will reach 32.935 billion yuan, with an annual compound growth rate (CAGR) expected to be about 32.93%.

Investment suggestion: paying attention to the update of the MoE model related to the MoE model may lead the new trend of the big model, while the development of AI products is expected to further enrich the overall ecology of its vertical field. Investors are advised to pay attention to the following related targets: Kunlun Wanwei, Jiecheng shares, Visual China, Zhongguang Natural selection, CITIC Publishing, Wanxing Technology, Insai Group, Blue cursor, Yuan Longyatu, Tianyu, Tom Cat and Central Plains Media, etc.

Risk hint: considering the uncertainty of policy, technological development and market acceptance, when considering investing in AI products and related targets, investors should pay close attention to factors such as policy uncertainty, the speed of AI technology development and market acceptance of AI applications, which may have an important impact on investment decisions.