Jakarta, INTI – Speculation is heating up online over the launch of DeepSeek-R2, the latest artificial intelligence (AI) model from Chinese startup DeepSeek. Touted as setting new standards in cost-efficiency and performance, the model is raising expectations amid the increasingly fierce technological rivalry between China and the United States.
DeepSeek-R2: An Efficient and Innovative Open-Source Model
DeepSeek, which previously captured global attention with the release of its V3 and R1 models at the end of 2024, is once again in the spotlight. DeepSeek-R2, the successor to R1—renowned for its reasoning capabilities—is reported to feature a hybrid mixture-of-experts (MoE) architecture with a total of 1.2 trillion parameters. This architecture allows tasks to be distributed across multiple smaller subnetworks, dramatically boosting computational efficiency.
Interestingly, the development of R2 is claimed to be 97.3 percent more cost-effective compared to the building of OpenAI’s GPT-4o. Leaks from the Jiuyangongshe platform indicate that the model was trained using Huawei’s Ascend 910B chips, which reportedly deliver up to 91 percent of the performance of Nvidia’s A100-based clusters.
Furthermore, R2 is expected to offer advanced multimodal capabilities, enabling it to process not only text but also images, audio, and even basic video comprehension—representing a significant upgrade over its predecessor, R1, which was primarily text-focused.
DeepSeek’s Distinct Strategy in the Global AI Race
Unlike many of its competitors, DeepSeek emphasizes resource efficiency and innovation in its training methods. DeepSeek-R2 leverages a technique called Generative Reward Modeling (GRM), allowing the model to learn human preferences without requiring massive human feedback datasets. Additionally, the Self-Principled Critique Tuning method enables R2 to self-evaluate and improve its responses based on internal principles, thereby enhancing accuracy and consistency independently.
DeepSeek’s strategy of rejecting large-scale investments to maintain research independence has also attracted attention. Rather than rushing to produce commercial products, the Hangzhou-based company focuses on long-term research, including its ambitious goal of developing Artificial General Intelligence (AGI).
Amid China's restricted access to Western technologies due to ongoing geopolitical tensions, DeepSeek-R2 stands as a powerful symbol of technological self-reliance. The use of domestic chips and local resources highlights China’s readiness to confront Western dominance in the tech sector, as noted by Deedy Das, Principal at Menlo Ventures, in a recent post on X.
Conclusion
The launch of DeepSeek-R2 marks a pivotal moment in the global AI landscape. It is not just about introducing a high-performance model but also about providing a tangible alternative to Western dominance in frontier AI development.
With its strong multilingual capabilities, integrated multimodal functions, and training efficiency innovations, DeepSeek-R2 has the potential to accelerate the democratization of AI technology.
If DeepSeek successfully meets these high expectations, the young company could emerge as a major force—not only transforming the AI market but also redefining how the world understands and builds artificial intelligence in the future.
Read More : 1.5 Billion Users and Counting: Google’s AI Overviews Is Reshaping How We Search