计算机行业分析报告：DeepSeek近期成果分析及V4影响力预测

Investment Rating - The report maintains an "Outperform" rating for the computer industry, expecting a relative increase of over 10% compared to the CSI 300 index in the next six months [1][50]. Core Insights - DeepSeek is set to launch its next-generation flagship AI model, DeepSeek V4, in mid-February 2026, which is anticipated to surpass the capabilities of Claude and GPT series models [3][11]. - The introduction of a new architecture independent of transformers in V4 is expected to mark a significant technological breakthrough, paving the way for advancements towards AGI [4][46]. - The report highlights the potential for cost reductions in model training, which could alleviate the current chip shortages in the domestic market [4][40]. - DeepSeek's commitment to an open-source approach is likely to enhance its competitive position against closed-source models from companies like OpenAI and Anthropic [41][45]. Summary by Sections 1. Latest Developments of DeepSeek - DeepSeek plans to release V4, which is positioned against the anticipated R2 model expected in May 2025 [11]. - The company has made significant advancements in adapting its models to domestic chips, enhancing compatibility and performance [12]. 2. Sparse Allocation Scheme - The introduction of the "Engram" memory module aims to improve model performance and decouple computation from memory constraints, addressing current GPU memory limitations [19][29]. - Experimental data indicates that allocating 20%-25% of sparse parameters to Engram optimizes overall model performance [22]. 3. Innovations in Information Transmission Architecture - The mHC architecture proposed by DeepSeek enhances information flow and stability in deep networks, improving training convergence speed by approximately 1.8 times [30][31]. 4. Long Text Input Compression - DeepSeek-OCR and its upgrade, DeepSeek-OCR2, utilize visual encoding to significantly reduce the number of tokens required for long text inputs, achieving high decoding accuracy even at substantial compression rates [34][37]. 5. Increased Transparency in Research - The update of the R1 paper from 22 to 86 pages reflects DeepSeek's commitment to transparency, detailing training processes and costs, which are significantly lower than those of leading models [39][40]. 6. Predictions for V4's Impact - V4 is expected to lower model costs and enhance performance, potentially transforming the competitive landscape of the AI industry [40][46]. - The model's deep integration with domestic chips is anticipated to support the development of the local computing ecosystem [47].