DeepSeek-V4：中國人工智慧硬體與軟體共同設計策略

隨著技術進步縮小了模型間的效能差距，中國開放原始碼模型透過軟體最佳化和擴展彌補硬體效能的不足，從而提升了其在全球的影響力。基於對「模型設計與系統工程」的根本性重構，中國模型即使在計算資源有限的情況下也能保持穩定的性能和卓越的成本績效。 DeepSeek-V4的出現便是這一趨勢的典型例證，它是首個在推理階段達到國內水平的中國模型，引領中國人工智慧產業進一步深化「軟硬體聯合設計」的發展路徑。

主要亮點

• 全球擴張：中國的開放原始碼模式正在全球擴張，透過軟體最佳化克服硬體限制。
• 高價值：透過系統重新設計，即使在運算能力有限的情況下，我們也保證了高效能和具成本效益。
• 生態系統轉型：向國內推理替代方案的轉變正在加速，協作硬體和軟體設計正在取得進展。

目錄

第1章：中國人工智慧的發展正從模型競爭轉向推理應用和開放原始碼生態系統的探索。

第2章：中國供應商的策略

第3章：TRI的觀點

As technical iteration narrows the capability gap between models, Chinese open-source models are expanding their global presence by compensating weaker hardware with software optimization and scale. Under compute constraints, Chinese models can still maintain solid performance and a strong price–performance advantage, based on a fundamental reshaping of “model design and system engineering.” The launch of DeepSeek‑V4, the first Chinese model to achieve domestic substitution in the inference stage, has become a flagship example, and is driving China’s AI industry to further deepen an integrated “AI hardware–software co‑design” development path.

Key Highlights

• Global Reach: Chinese open-source models expand globally, overcoming hardware limits via software optimization.
• High Value: System redesigns ensure strong performance and cost-efficiency despite compute constraints.
• Ecosystem Shift: Domestic inference substitution accelerates hardware-software co-design.

Table of Contents

1. China’s Advancement in AI to Shift from Battles of Models to Exploration of Inference Applications and Open-Source Ecosystems

• Figure 1: Model Invocation in April 2026
• Table 1: Summary of Low-Precision Models
• Figure 2: China’s Countermeasures towards Restraints in Computing Power
• Figure 3: Key Ugrades to DeepSeek-V4
• Table 2: Comparison between Checkpointing and Compute While Transmitting
• Table 3: Comparison between CUDA and CANN

2. Strategies among Chinese Vendors

• Figure 4: China’s AI Software-Hardware Ecosystems
• Figure 5: China’s AI Hardware-Software Integration
• Table 4: Comparison of DeepSeek-V3 MLA and DeepSeek-V4 MQA Architectures

3.TRI’s View