DeepSek-R2 Launched At The End Of August, China Is More Independent In AI

DeepSek-R2, is scheduled for release between August 15 and 30, 2025. (photo: x @soltomm_ ·)

JAKARTA - China is increasingly aggressive in its efforts to achieve independence in the field of artificial intelligence (AI). This is marked by preparations for the launch of the next generation of large language model (LLM), DeepSek-R2, which is scheduled between 15 and 30 August 2025.

This schedule is revealed from the interaction with the DeepSek AI model itself, signaling a very close launch, just days after the ChatGPT-5 was released.

DeepSek-R2 is expected to bring significant jumps in architecture. The model will adopt a more sophisticated Mixture of Experts (MoE) structure and integrate smarter gaming networks to handle heavy inference workloads.

According to sources familiar with its development, this model can reach a scale of up to 1.2 trillion parameters, almost double that of DeepSek-R1 which has 671 billion parameters. This figure is still lower than the ChatGPT-4/5 which has more than 1.8 trillion parameters.

Independent From West Hardware

As a strategic step towards self-sufficiency in AI, DeepSeek-R2 is fully trained using Huawei's Ascend 910B chip. Huawei's computing cluster, which generates FP16 performance of 512 PFLOPS with 82% utility, is reportedly capable of reaching 91% of Nvidia A100 cluster performance. Analysts see this as a crucial step to reduce China's dependence on US-made AI hardware.

DeepSek-R2 training reportedly cost 97% lower than GPT-4, thanks to the use of local hardware and optimization techniques. Analysts expect DeepSek will offer API access at a lower price, potentially shaking prices models currently dominated by OpenAI and Anthropic.

Anticipate the launch of DeepSeek-R2 has triggered movement in China's technology stock market. chipmaker AI, Cambricon, experienced a 20% increase in share price, pushing its market capitalization past 355 billion yuan (around Rp801.2 trillion).

In related news, Huawei also launched Unified Cache Manager (UCM), a new AI inference framework. Designed to accelerate model inference, UCM optimizes the way KV Cache data is handled across memory levels such as HBM, DRAM, and SSD.

In testing with China UnionPay, Huawei reported a drop in latency of up to 90% and a 22x increase in throughput. The company plans to make UCM an open source in September.

Overall, the launch of DeepSek-R2 and the introduction of UCM framework from Huawei signaled a major shift in China's AI ambitions. This development leads to a future where China can build and run high-performance AI systems without relying on chips or software from Western countries.