There has been much consternation in the United States over the past few weeks around the supposed proliferation of open-weight AI models from China. Even so, DeepSeek's newly instituted peak-hour pricing mechanism belies those fears, and suggests that most of its users still reside in Asia.
While this paradigm is also likely to hold true for some of China's other high-flying AI models, such as Alibaba's Qwen series of LLMs, the sheer scale of oncoming demand is truly a marvel to behold, with Qwen models alone occupying a download footprint on Hugging Face that is 2.6x that of Meta's!
Towards the end of July, DeepSeek launched a refreshed version of its latest Flash-class model, dubbed the V4-Flash-0731. Even though the model has just 284 billion parameters, it offers a performance that is similar to Anthropic's Opus 4.8, which is widely believed to span multi-trillion parameters! What's more, 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.
Then, just last week, DeepSeek started rolling out the V4-Pro on its API and Chat, pricing the model at $0.435 per 1 million tokens of input and $0.87 per 1 million tokens of output. As per the preliminary benchmarks on WeChat, the V4-Pro model outcompetes Opus 4.8 on Terminal Bench 2.1, Cybergym, DeepSWE, and AutomationBench benchmarks
Do note that DeepSeek's limited compute capacity has likely been overwhelmed by the oncoming demand for its V4-class models. After all, the AI lab was second only to Anthropic in terms of token volume in July, and might even clinch the apex spot in the coming months.
Consequently, DeepSeek has now implemented a peak-hour pricing mechanism :
- Input price for V4-Flash rises from $0.14 per 1 million tokens to $0.22/$0.44 (off-peak/peak)
- Output price for V4-Flash rises from $0.28 per 1 million tokens to $0.66/$1.32 (off-peak/peak)
- Input price for V4-Pro rises from $0.435 per 1 million tokens to $0.66/$1.32 (off-peak/peak)
- Output price for V4-Pro rises from $0.87 per 1 million tokens to $1.98/$3.96 (off-peak/peak)
Critically, DeepSeek's peak hours run from 09:00 - 12:00 and 14:00 - 16:00 Beijing time, implying that most of its users still reside in Asia, placating worries around US enterprises turning en masse towards these models.
Alibaba has launched two major Qwen 3.8-class models recently. First up, the Qwen 3.8-Max (2.8 trillion parameters) is second only to Moonshot's Kimi K3 in terms of performance. More intriguing, however, is the Qwen 3.8 (27 billion parameters) model, which sports coding capabilities that are similar to Opus 4.5, and yet can run on a single MacBook !
Given the fact that many of China's open-weight models now sport capabilities that are roughly equal to those offered by the proprietary models from OpenAI and Anthropic, it is hardly a surprise that they are surging in popularity.
In fact, Hugging Face - one of the largest repositories of open-weight AI models - noted just last week that Alibaba's Qwen models now account for 151,448 derivatives (total downloads), which is roughly 2.6x the download count for Meta's open-weight AI models!
Follow Wccftech on Google to get more of our news coverage in your feeds.