Last time it was during the Spring Festival holiday, could this time be during the May Day holiday? That's also good, as every holiday brings beautiful expectations.
Recently, various news sources revealed that DeepSeek is about to launch the next generation of large models, DeepSeek R2.
According to social media revelations, the core highlights are as follows:
Architectural innovation: Utilizing self-developed Hybrid MoE 3.0 architecture, with dynamic activation parameters reaching 1.2 trillion and actual computational consumption only 78 billion, resulting in significant efficiency improvements.
Hardware localization: Trained based on Huawei Ascend 910B chip cluster, achieving a computing power utilization rate of 82%, performance close to Nvidia A100 clusters' 91%;
Multimodal leap: Achieved an accuracy of 92.4% in the COCO image segmentation task, surpassing the CLIP model by 11.6 percentage points.
Vertical field landing: Medical diagnosis accuracy exceeds 98%, and the false positive rate in industrial quality inspection is reduced to 7.2 parts per million, with the practicality of technology reaching new heights.
It is worth noting that three months ago, the release of DeepSeek R1 caused Nvidia's market value to evaporate by $600 billion in a single day, while R2's 'low cost + high performance' combination will undoubtedly bring a more severe impact on American tech giants that rely on high-premium chips.
Technological breakthrough
The parameter scale of the DeepSeek R2 large model has been reported to reach an astonishing 1.2 trillion, nearly doubling the 671 billion parameters of the previous generation R1. This number is close to the level of top international models such as GPT-4 Turbo and Google Gemini 2.0 Pro. The explosive growth in parameter count indicates a significant enhancement in the model's learning and complex task processing capabilities.
DeepSeek R2 adopts a mixture of experts (MoE) architecture, a technology that assigns tasks to multiple 'small expert' modules. In simple terms, the model automatically selects the most suitable 'expert' to handle different tasks, improving efficiency while reducing waste of computational resources.
According to the revelations, the dynamic activation parameter of R2 is 78 billion, and the actual computational consumption is only 6.5% of the total parameters. This design allows the model to maintain high performance while significantly reducing operating costs.
In terms of training data, DeepSeek R2 uses a high-quality corpus of 5.2PB (1PB = 1 million GB) covering fields such as finance, law, and patents. Through multi-stage semantic distillation technology, the model's instruction-following accuracy has been improved to 89.7%. This means it is better at understanding complex human instructions, such as analyzing legal documents or generating financial reports.
Cost reduction of 97.3%
The biggest breakthrough of DeepSeek R2 is still the significant reduction in costs. According to revelations, its unit inference cost is 97.3% lower than that of GPT-4. For example, generating a 5000-word article with GPT-4 costs about $1.35, while DeepSeek R2 only costs $0.035.
The core reason for the cost reduction lies in the optimization of hardware adaptation. DeepSeek R2 is trained based on the Huawei Ascend 910B chip cluster, achieving a chip utilization rate of 82%. In comparison, the performance of similar Nvidia A100 clusters is 91%.
This indicates that domestic chips in the AI training field have approached international leading levels and may even be able to break free from dependence on Nvidia.
Multimodal capability
Another major highlight of DeepSeek R2 is the enhancement of multimodal capabilities. In the visual understanding module, it adopts a ViT-Transformer hybrid architecture, achieving an accuracy rate of 92.4% in the COCO dataset object segmentation task, an improvement of 11.6 percentage points over the traditional CLIP model.
In simple terms, it can more accurately identify objects in images, such as distinguishing pedestrians, vehicles, and traffic signs from a street scene photo.
In addition, R2 supports 8-bit quantization compression, reducing the model size by 83% with a precision loss of less than 2%. This means that in the future, smartphones and smart home devices can also run high-performance AI locally without relying on cloud servers.
AI Global Game
The revelations about DeepSeek R2 have triggered a strong reaction in the capital market. Due to cost advantages and technological autonomy, it may pose a threat to American tech companies that rely on Nvidia GPUs.
Analysts predict that if R2's performance is accurate, Nvidia's stock price may face short-term fluctuations, while Chinese AI industry chain-related companies may welcome a new round of growth.
This event also reflects the new pattern of the global AI competition. DeepSeek R2 proves that breakthroughs can also be achieved through architectural innovation and domestic hardware adaptation. The utilization data of Huawei Ascend chips (82%) indicates that domestic computing power infrastructure has international competitiveness.
Although the revelations are exciting, some industry insiders have pointed out contradictions in the information. Discussions on external networks have shown that Chinese informal channel messages are being translated and disseminated, further increasing uncertainty.
DeepSeek has not yet confirmed the release date, but combined with ZZJ's recent collective learning trends in artificial intelligence, the dual benefits of policy support and technological breakthroughs may accelerate R2's market entry.
From chips to algorithms, from data to applications, every link in the Chinese AI industry chain is accelerating autonomy. Huawei Ascend replaces Nvidia, and a 5.2PB Chinese corpus builds vertical barriers—these actions are part of a life-and-death race concerning technological discourse power for the next decade.
