DeepSeek's New Flash Model Tops OpenRouter With 33.6 Trillion Weekly Tokens, Up 801% in a Week

The most-used AI model on OpenRouter, one of the world's largest LLM routing platforms, now has a name and a growth curve that are hard to ignore. Data for the week of October 19-25, published by Chinese outlet Daily Economic News based on OpenRouter figures, shows DeepSeek's V4.1-Flash-1017 processed 33.6 trillion tokens in a single week — up 801% from the week before — putting it at the top of the global usage rankings.

The migration behind that number is visible in the runner-up slot: the previous DeepSeek V4.1 Flash, launched September 10, fell 45% to 18 trillion weekly tokens as developers moved to the new checkpoint. A two-week-old model did not just catch the old one; it absorbed most of its traffic and added more on top. OpenRouter's live leaderboard, computed through the October 26 usage bucket, still shows the same picture: DeepSeek V4.1-Flash-1017 leads at 35.4 trillion trailing-week tokens, more than double the second-place variant.

The broader weekly totals put the scale in context. Global usage across all models on the platform reached 188 trillion tokens for the week, up 14.63% week over week. Models from Chinese developers processed 93.54 trillion of those tokens, up 18.14%, while U.S.-origin models processed 22.47 trillion, up 29.44% from a smaller base. It is the 26th consecutive week Chinese models have topped the platform's usage charts — and for the first time in these weekly rankings, all five of the top slots went to Chinese models.

Usage share tells a similar story from a different angle. DeepSeek accounted for 26.5% of all text requests on the platform during the week beginning October 19, ahead of Google at 20.1% and OpenAI at 18.5%, according to OpenRouter's market-share breakdown. On the requests leaderboard by individual model, free and cheap Chinese flash models occupy most of the top spots, which is exactly the profile you would expect from a developer ecosystem optimizing for cost per token.

Caveats belong in any reading of this data. OpenRouter is one routing platform, skewed toward developers and open-weight models, and it excludes private requests; it does not measure total global inference across all providers and APIs. Token counts also reward verbose models. A model can dominate routed traffic while remaining a rounding error in consumer chatbots, and OpenRouter itself warns that higher token totals show usage, not capability.

Still, the direction of travel matches everything else DeepSeek has shipped this month: the V4.1-Flash-1017 checkpoint arrived with an expanded in-house search API in late October, the DeepSeek Harness coding agent has been climbing OpenRouter's app rankings, and the company is reportedly raising a record funding round of up to 100 billion yuan. Cheap, fast, open-weight models are where agentic workloads want to live — and right now, on the platform where developers actually route their agents, DeepSeek owns that lane.