On November 12, 2026, the Baidu World 2026 conference opened in Beijing, where Baidu founder Robin Li officially launched "Ciyuansuo" (Token Exchange), a large-model trading platform positioned as a neutral and trustworthy exchange for large language models. One month after going live, the platform already hosts more than 100 supply resources, with daily calls exceeding 50 million and daily token usage above 700 billion. The problem Baidu wants to solve is straightforward: models keep multiplying, yet enterprises are growing afraid to buy one.
Verification and Credit Scoring: Check Identity on Entry, Keep Scoring Performance
The platform rests on two capabilities. Verification means checking the credentials of listed models and providers, then confirming the authenticity of resources through model fingerprinting and capability, performance, and availability testing — keeping swapped models and inflated specs out. Credit scoring means rating providers on service quality, fulfillment, security compliance, and commercial performance, alongside regular spot checks and full-link real-time monitoring of calls, with users alerted when a credit score falls. As Li put it, Ciyuansuo should let enterprises and individuals easily find, buy, and use trustworthy model tokens, just as a stock exchange helps investors discover companies.
One Month of Data: Real Demand, Still Early Scale
The numbers deserve a closer look. In the first nine months of this year, new models appeared almost daily, sometimes several in a day, and demand for model calls surged with them; yet the token relay stations on the market take no responsibility for model quality or service stability, and the same model can vary widely in price and quality across platforms. Ciyuansuo's 700 billion daily tokens after one month show that demand for platform-based metering is real, but set against the roughly 500 trillion tokens consumed daily across China in the CAICT report (June 2026 figures), its volume is still early-stage. The platform already lists 42 mainstream models, including DeepSeek, Kimi, and Zhipu GLM (per media reports on November 5); the domestic site is live, and the overseas site is coming soon.
For enterprise buyers, the value of Ciyuansuo is turning "is this model real, and will its promises hold" into a checkable score, instead of re-testing on every purchase. For model providers, credit ratings could become a new competitive dimension: stable performers win more orders, while shrinking service gets publicly flagged. But one structural issue cannot be avoided: Baidu is itself a model provider, so the platform's neutrality will need continuous proof. Full trade matching, price disclosure, and settlement mechanisms have not yet been spelled out either — those depend on what lands next.
The same conference also unveiled the ERNIE 5.5 Flash Preview, two ERNIE application models, and the Tianchi 1,024-card supernode. For Ciyuansuo, the next things to watch are how the overseas site performs after launch, and whether the market responds the first time a credit rating genuinely affects a provider's orders.
