Chinese and Open-Weight Models' Share of OpenRouter Tokens
This chart follows two shares of the tokens processed on OpenRouter, a marketplace that routes developers' and apps' requests to hundreds of AI models: the share on models from Chinese developers, and the share on models whose weights are public. Both are weekly, from January 2025, and count only tokens on models whose developer is known.
What does it show?
Chinese developers' models carried 70% of OpenRouter's tokens in the week to September 20, 2026, up from 7% in the first week of 2025; open-weight models carried 75%.
Methodology
OpenRouter Datasets API, rankings-daily: tokens (prompt plus completion) for the day's 50 most-used models on OpenRouter, from January 1, 2025, summed into Monday-to-Sunday weeks, each dated by its Sunday; a week needs at least five days of data and the current, unfinished week is left out. The developer is the model id's prefix, and a model counts where its publisher is based (a Western fine-tune of a Chinese model counts as Western). The denominator is tokens on top-50 models with a known developer: the row for all models outside the day's top 50, stealth releases, OpenRouter's own routers and any developer not yet mapped are left out (10% of all tokens over the period). Open-weight means the model's weights are public: OpenRouter's model list links it to Hugging Face, or, for retired models it no longer lists, the developer releases only open or only closed models, or a checked override; tokens with no evidence either way (0.0%) are left out of that share. One marketplace, not the AI market: developers and apps, heavy on coding tools and roleplay, and none of ChatGPT, Claude.ai or company traffic that goes straight to the labs. Each lab counts tokens with its own tokenizer, so shares are approximate. Source: OpenRouter (openrouter.ai/rankings), as of 2026-09-27; CC BY 4.0.