In the past few weeks, there has been significant concern in the United States regarding the proliferation of open-source AI models from China. However, recent moves by DeepSeek, a prominent AI lab, suggest that most of its users are still based in Asia, specifically in East Asia.
DeepSeek recently launched an updated version of its Flash-class model, the V4-Flash-0731, with 284 billion parameters. Despite its relatively modest parameter count, the model boasts performance comparable to Anthropic’s Opus 4.8, which is believed to have a parameter count in the multi-trillion range. DeepSeek priced the V4 Flash 0731 at just $0.14 per 1 million tokens of input and $0.28 per 1 million tokens of output.
Just last week, DeepSeek introduced the V4-Pro, further boosting its API and chat offerings. The V4-Pro is priced at $0.435 per 1 million tokens of input and $0.87 per 1 million tokens of output. Preliminary benchmarks from WeChat indicate that the V4-Pro outperforms Anthropic’s Opus 4.8 on Terminal Bench 2.1, Cybergym, DeepSWE, and AutomationBench.
To manage the overwhelming demand, DeepSeek has implemented a peak-hour pricing mechanism. The peak hours are set between 09:00 and 12:00 and 14:00 and 16:00 Beijing time. This pricing strategy highlights that most of DeepSeek’s users are still based in East Asia, particularly in China. This move alleviates concerns in the US about enterprises turning to Chinese models for their AI needs.
Alibaba, another Chinese AI powerhouse, has also been making waves with its Qwen models. Alibaba recently launched two major Qwen 3.8-class models: the Qwen 3.8-Max (2.8 trillion parameters), which is second only to Moonshot’s Kimi K3 in performance, and the Qwen 3.8 (27 billion parameters), which offers coding capabilities similar to Anthropic’s Opus 4.5 and can run on a single MacBook.
These capabilities, which rival those of proprietary models from OpenAI and Anthropic, have contributed to the surge in popularity of Chinese AI models. According to Hugging Face, the largest repository of open-source AI models, Alibaba’s Qwen models have already garnered 151,448 derivatives (total downloads), approximately 2.6 times the number of downloads for Meta’s open-source AI models.
In conclusion, the strategic pricing and operational moves by DeepSeek and Alibaba indicate that the bulk of users for these models remain in East Asia, particularly China. This reality serves to calm US fears about a potential Chinese AI takeover, while highlighting the competitive edge of Chinese models in the AI landscape.
Thinking about building an AI product?
Get in Touch