Developers on OpenRouter moved to DeepSeek's October 17 release within days, cutting its predecessor's usage by 45%, as DeepSeek reportedly nears a RMB100 billion funding round.
DeepSeek’s latest Flash update became the most used model on OpenRouter in its first full week there. DeepSeek V4.1-Flash-1017 processed 33.6 trillion tokens, the units of text that models read and write, between October 19 and 25, up 801% on the previous week. The figures come from calculations by Daily Economic News (每日经济新闻) based on OpenRouter data, republished by the 21st Century Business Herald, and OpenRouter’s own weekly rankings show the same 33.6 trillion and 801%.
The same tally put Chinese models at 93.54 trillion tokens for the week, up 18.14%, and US models at 22.47 trillion, up 29.44%, out of a platform total of 188 trillion, up 14.63%. By that count, Chinese models have handled more weekly usage than US models for 26 weeks in a row, and the five most used models of the week were all Chinese.
OpenRouter is one marketplace, a service that routes developers’ requests to hundreds of models through a single API, and its numbers cover only that channel. Jiemian pointed out that the data count public traffic through OpenRouter’s aggregated API and leave out requests made through companies’ own websites, apps and private deployments, though it called the rankings a direct read on demand from independent developers and small businesses.
Developers moved quickly from DeepSeek’s old Flash to the new one
DeepSeek released V4.1-Flash-1017 on October 17 as an update to the existing V4.1 Flash model, with improvements in reasoning, coding, agent tasks and visual understanding, TechNode reported. Its weights are on Hugging Face under the MIT license. The same release added paid Search and Fetch APIs for web retrieval. Because the version went live near the end of the previous reporting week, the 801% comparison starts from a partial week of usage.
The earlier V4.1 Flash, launched on September 10, ranked second with 18 trillion tokens, down 45%. Jiemian read the decline as developers migrating rapidly to the new version. DeepSeek’s two Flash versions together processed 51.6 trillion tokens, more than a quarter of the platform’s weekly total. Measured by requests, a different count from the token totals, OpenRouter’s market-share chart by model author for the week beginning October 19 put DeepSeek at 26.5%, ahead of Google at 20.1% and OpenAI at 18.5%.
Below DeepSeek, MiniMax’s M3.1 Flash, listed in a free version, entered at third with 13.7 trillion tokens in its first appearance, after its October 19 release. MiniMax says the model runs inference nearly three times faster and costs about 70% less to use, according to Jiemian. Z.ai’s GLM 5.3 Flash was fourth with 13.2 trillion, up 8%, and Tencent’s Hy4 preview fifth with 8.97 trillion, up 19%, its fourth straight week in that spot.
Low-cost Flash models carry most of the volume
The first non-Chinese entries appear at sixth and seventh place. OpenRouter shows OpenAI’s GPT-6 Luna at 8.35 trillion tokens, up 35%, and Anthropic’s Claude Opus 5.5 at 8.27 trillion, up 34%. Xiaomi’s MiMo-V2.6-Flash fell 39% to 6.07 trillion, Nvidia’s free Nemotron 3 Ultra processed 5.85 trillion, and Z.ai’s full GLM 5.3 closed the top ten with 5.05 trillion.
Jiemian attributed the strength of Chinese Flash models to three factors: architectures that lower the cost of each inference run, updates released on a near-weekly cycle, and new versions that keep compatibility with the old API without raising prices, which makes switching easy. OpenRouter’s model page lists V4.1-Flash-1017 at US$0.04 per million input tokens and US$0.50 per million output tokens.
The country totals need care. They are Daily Economic News calculations rather than figures OpenRouter publishes, and the reports do not say how models from other countries were counted. US usage also grew faster than Chinese usage this week, even as China’s lead in absolute volume widened.
A funding round and a full week of data come next
The usage figures land as DeepSeek negotiates new capital. The Shanghai Securities News reported on October 22, according to Jiemian, that DeepSeek planned to raise about RMB100 billion with Tencent and CATL as the main investors, a round EastFrontier reported was nearing completion. Jiemian also reported that DeepSeek released DeepInfer, its own large-scale inference framework, as open source on October 20, and said the software supports high-volume serving of the V4.1 Flash series.
The week of October 26 to November 1 will be the first in which V4.1-Flash-1017 is measured against a full week of its own usage, so its growth rate will drop sharply even if volume holds. Two numbers will say more about demand than the headline percentage: whether DeepSeek’s combined Flash total stays above 50 trillion tokens, and whether MiniMax’s M3.1 Flash keeps its third place once the free listing is no longer the main way developers try it.