Here are the latest updates about DeepSeek from China in the past month (August 23, 2025, to September 23, 2025), presented clearly and simply:
- New Model Release (V3.1-Terminus): DeepSeek updated its online AI model to version V3.1-Terminus, announced on September 22, 2025. This update aims to improve performance, but specific details about changes were not provided.
- Training Cost Revelation: On September 19, 2025, DeepSeek revealed it spent only $294,000 to train its R1 model, significantly less than U.S. rivals like OpenAI, which spent over $100 million on models like GPT-4. The R1 model used 512 Nvidia H800 chips, designed for the Chinese market due to U.S. export restrictions. This low-cost approach has sparked debate about China’s AI development capabilities.
- Huawei Collaboration for Safe Model: On September 19, 2025, Huawei announced it co-developed a safety-focused version of DeepSeek’s R1 model, called DeepSeek-R1-Safe, using 1,000 Huawei Ascend AI chips. This version is designed to prevent discussion of politically sensitive topics, achieving near 100% success in blocking harmful content, though it struggles (40% success) with disguised prompts. DeepSeek itself was not directly involved in this project.
- Huawei Chip Adoption: DeepSeek has started using Huawei’s Ascend chips for model training, alongside Nvidia chips, as reported on August 29, 2025. This shift was encouraged by Chinese authorities to reduce reliance on foreign technology, despite challenges like stability issues and slower inter-chip connectivity with Huawei chips.
- New AI Model Development: DeepSeek is working on a new AI model designed to perform multi-step tasks with minimal user direction, as reported on September 4, 2025. This model aims to enhance automation, but no release date was confirmed.
- Unique Data Format (UE8M0): On August 24, 2025, DeepSeek introduced a radical FP8 data format (UE8M0) for AI processing, which prioritizes range over precision to simplify hardware needs. This move supports China’s strategy to build an integrated AI-semiconductor ecosystem with domestic chipmakers like Huawei and Cambricon.
- Security Concerns Raised: On September 17, 2025, research indicated DeepSeek’s AI provides less secure code for groups disfavored by the Chinese government, like Falun Gong, raising concerns about bias and security in its outputs.
DeepSeek is not significantly hindered in moving forward due to a lack of Nvidia chips, though challenges exist. Recent updates show DeepSeek is adapting by using Huawei’s Ascend AI chips alongside Nvidia’s H800 chips, as reported on August 29, 2025. This shift, encouraged by Chinese authorities to reduce reliance on foreign technology, allows DeepSeek to continue training models like R1, which was developed using 512 H800 chips for only $294,000. However, Huawei’s chips face issues like stability and slower inter-chip connectivity compared to Nvidia’s, which could impact efficiency. Despite these challenges, DeepSeek’s progress—such as releasing V3.1-Terminus and developing new models—indicates it is finding workarounds and not stalled by chip limitations.