This picture illustration shows a mobile phone's screen showing the
This picture illustration shows a mobile phone's screen showing the logo of Chinese AI MiniMax.

Chinese large language models recorded higher usage than US models on OpenRouter for the 23rd consecutive week, although their weekly usage has declined for two weeks in a row. The anonymous Space Bunny model topped the single-model rankings, but its usage is not counted under either China or the United States.

Calculations by the Shanghai-based National Business Daily, based on OpenRouter data, showed that models listed on the platform generated 166 trillion tokens in total usage last week, from September 28 to October 4. Chinese models accounted for 57.46 trillion tokens, compared with 16.2 trillion for US models — about 3.5 times as much. The figure for Chinese models fell about 7.6 per cent from 62.22 trillion tokens the previous week, which had already marked a week-on-week decline of 7.77 per cent.

Among individual models, four of the global top five were from China. DeepSeek V4.1 Flash ranked second with 256 trillion tokens, followed by Zhipu GLM 5.3 Flash. Xiaomi’s MiMo-V2.6-Flash and Tencent’s Hy4 preview ranked fourth and fifth respectively. The top spot went to the anonymous Space Bunny Alpha, or “Space Bunny”, which recorded 387 trillion tokens, up 179 per cent week on week.

Under National Business Daily’s previous methodology, anonymous models are not assigned to any country. The same applies this week, as Space Bunny’s 38.7 trillion tokens alone, when added to DeepSeek’s 25.6 trillion, exceed China’s total of 57.46 trillion tokens.

What is Space Bunny?

Space Bunny was launched on OpenRouter on September 23, promoting fast reasoning and coding capabilities. It supports a 1-million-token context window and can process text, image and video inputs. OpenRouter classifies it as a “stealth” model and offered it free for a limited time, with the listing marked for removal on October 5. The free access alone may have driven up usage; usage figures indicate popularity, not necessarily model quality.

As for its developer, independent tests have pointed to MiniMax. A third party compared token counts across 50 test strings and found results consistent with the MiniMax series; another set of 24 tests reached a similar conclusion. MiniMax launched M3.1-Flash-Preview through its MiniMax Code platform on September 27, with specifications similar to Space Bunny. However, neither side has officially confirmed any relationship between the two. MiniMax, which is listed in Hong Kong under stock code 0100, fell about 10 per cent after the market opened on September 28, closing at HK$247.40.

Why Flash models are gaining popularity

Developers are increasingly turning to faster and cheaper models for routine tasks such as code completion and text generation. Reports said DeepSeek V4 Flash charges about US$0.14 per million input tokens, compared with US$5 for OpenAI’s GPT-5.5 at its June pricing. OpenRouter data staff have also told CNBC that Chinese open-source models can be 60 to 90 per cent cheaper than leading US closed-source models. Flash models typically trade a lower number of activated parameters for speed and lower costs. Xiaomi’s MiMo-V2.6-Flash, for example, has 309 billion total parameters, of which 15 billion are activated.

How to interpret this “first”

  • Platform characteristics: OpenRouter is a model API aggregation platform used mainly by developers and start-ups. AI assistants deployed through private enterprise systems or embedded in applications such as Douyin and WeChat are not included in the figures.
  • Reasons for the lead: Chinese models’ lead is mainly linked to pricing and usage in agent workflows, rather than being driven solely by Chinese developers. An OpenRouter operations director previously said Chinese models accounted for a relatively high share of agent workflows at US companies.
  • Narrowing gap: The ratio of Chinese to US weekly usage, as calculated by National Business Daily, reached 14.1 times in late July. It is now about 3.5 times.
  • Prices and revenue: US models command higher prices, so the revenue gap does not correspond directly to the usage gap.

What to watch next

  1. Whether MiniMax will formally launch a similar model after Space Bunny is removed, and whether usage can be sustained once fees are introduced.
  2. Whether the price war in the Flash segment will continue, and which providers can maintain a balance between cost and performance.
  3. Whether Chinese models can move beyond platforms and enter the procurement lists of European and US companies. Language and data-compliance requirements remain major obstacles.