CITIC SEC has released a research report stating that the DeepSeek-V4 preview version has doubled its parameter count compared to the previous generation while matching the performance of leading global closed-source models. The model achieves state-of-the-art results among open-source alternatives with continued optimization of computing costs, ushering in an era of cost-effective million-token context models. DeepSeek-V4 features numerous innovations including upgrades to hybrid attention mechanisms, mHC, and Muon architectures. Domestic computing power and domestic models continue to align and deeply adapt, marking the arrival of a golden development period for domestic computing infrastructure.
DeepSeek-V4 maintains its open-source strategy with significantly reduced costs while enhancing capabilities in context length and agent functionality, comprehensively benefiting complex application scenarios. Investment strategy recommendations focus on three key themes: 1) AI infrastructure: DeepSeek's deep adaptation with domestic computing power demonstrates mutual advancement between domestic chips and models. 2) AI applications: The continued open-source approach dramatically reduces input/output costs while improving context length and agent capabilities, benefiting complex application scenarios and companies with competitive barriers. 3) Model developers: DeepSeek's new generation models are expected to collaborate with other domestic models to accelerate China's AI global expansion, while further reducing training and inference costs. More affordable tokens will drive increased global API usage for large models.
Key observations from CITIC SEC include: The DeepSeek-V4 preview version, launched on April 24, doubles parameter count from the previous generation while introducing cost-effective million-token context models. The release includes two base models: DeepSeek-V4-Pro positioned as a high-performance expert model with 1.6T total parameters (49B activated), and DeepSeek-V4-Flash as a cost-efficient rapid model with 284B total parameters (13B activated). Following official discounts announced on April 25, pricing for DeepSeek-V4-Pro dropped to ¥3/MTokens input and ¥6/MTokens output, representing extremely high cost-effectiveness compared to global leading models.
Performance evaluations show DeepSeek-V4-Pro-Max outperforming open-source models in knowledge tasks while narrowing the gap with closed-source alternatives. In reasoning tasks, it surpasses GPT-5.2 and Gemini-3.0-Pro, slightly trailing GPT-5.4 and Gemini 3.1-Pro. The Flash version matches GPT-5.2 and Gemini-3.0-Pro. In agent tasks, the Pro version competes with leading open-source models while slightly lagging cutting-edge closed-source alternatives. Industry testing has demonstrated practical long-context capabilities with improved stability and significant programming advancements, ranking third among open-source models on Arena.ai's coding platform.
Technical innovations include hybrid attention mechanisms combining CSA and HCA architectures to compress computational overhead and cache usage. The mHC structure updates residual connection paradigms while introducing online hybrid distillation strategies during post-training. Computational optimizations feature breakthroughs in computing-to-communication ratios, heterogeneous KV Cache, and FP4 quantization awareness. The computing-communication balance theory represents a major advancement for MoE model optimization, establishing a golden ratio of 6144 FLOPs/Byte that benefits domestic hardware development.
Domestic computing power continues to align with model development, with domestic chips announcing day-zero adaptation support. This reinforces the certainty of domestic AI chip adoption, transforms demand structures toward inference cards and super-nodes, and raises the commercial ceiling for domestic computing solutions. The open-source strategy with reduced costs and enhanced capabilities benefits complex application deployment, particularly for software companies with industry expertise, embedded enterprise systems, and specialized data advantages.
Risk factors include potential delays in AI technology development and application expansion, insufficient cost reduction in computing power, misuse of AI causing social impacts, data security concerns, information security risks, and intensifying industry competition.
Comments