Baidu Launches a "Stock Exchange" for AI Models, Listing Rivals Like DeepSeek and Kimi and Credit-Rating Providers to Stop Model Swapping

Baidu wants to be the stock exchange of AI models. On Nov. 12, Baidu founder Robin Li unveiled Ciyuansuo, roughly "Token Exchange," at Baidu World 2026: a marketplace that gathers AI models and the companies that serve them, rates them, and lets businesses and individuals buy access with more confidence. It doesn't just sell Baidu's own ERNIE models. Rivals such as DeepSeek, Kimi and Zhipu's GLM are on the shelf too.

A few terms

Token: the unit AI models use to process text, roughly a word or part of one. Using a model is generally billed per token, much like electricity is billed per kilowatt-hour.

Model provider: the same open model, DeepSeek for example, can be hosted by many companies on their own servers and sold as a service, each with different prices, speed and reliability.

Model fingerprinting: running a specific set of tests to "identify" a model and check whether a provider is actually serving the model it claims.

Why build it

According to IT Home's on-site report, Li said that when developers or businesses actually try to put models to work, more choice makes picking harder: the same model is expensive in one place and cheap in another, sources are all over the map and quality varies, so people end up unsure and afraid to commit. According to PANews, citing STAR Market Daily, Li also said that in the first nine months of this year more than one new model was released almost every day, and the surge in demand spawned a crop of token resellers.

The trouble with resellers is that buyers can't easily tell whether they're getting the real thing. A model may be quietly swapped for a cheaper, smaller one, or its specs may be overstated.

How it works

Li compared it to a stock exchange: an exchange attracts good companies to list so investors can find them, and Token Exchange aims to gather good models and providers so businesses and individuals can easily find and buy trustworthy models.

IT Home reports that the exchange runs entry tests on models it lists, including fingerprint comparison, capability tests, performance metrics and availability checks, and keeps monitoring quality afterward, to keep model swapping and inflated specs out. It also scores providers on service quality, delivery on promises, security and compliance, and contract performance, and assigns credit-rating labels. Users are alerted when a provider's rating drops, and all of the data goes into publicly searchable credit records.

Baidu's figures: one month after launch, the exchange has more than 100 supply sources, more than 50 million calls a day (per PANews) and more than 700 billion tokens processed a day.

The platform didn't appear out of nowhere today. 36Kr reported on Nov. 5 that it had already gone live quietly, with 42 models at the time and several providers per model, so users could compare prices, time to first token, throughput and seven-day success rates side by side. What happened today is Li's official launch on stage, along with usage figures.

Why it matters

This marks a change in Baidu's role. It used to sell its own ERNIE models mainly through its own cloud. Now it puts ERNIE on the same shelf as competitors' models to be compared on price, acting more like a referee and a marketplace. 36Kr noted that Alibaba's Bailian and ByteDance's Volcano Ark are also aggregating many models, and that Baidu is competing for developers' entry point and for the new business of trading model access.

That also raises an obvious question: Baidu is both a player and the referee. Whether its ratings of rivals are neutral will take time and public data to prove. 36Kr also pointed out that integrating many third-party providers, unifying APIs and billing, and continuously monitoring reliability are hard, and that any skew in the ratings would directly hurt developers' trust.

What it means for you

You won't buy tokens on the exchange yourself, but many AI apps you use pay model providers by the token behind the scenes. If buying model access works like buying stock, with public prices, report cards and credit histories, app makers can more easily find models that are cheap and reliable, and swapped-out models or degraded service may become less common. That could eventually show up as cheaper, steadier AI products. It's also a sign that AI models are turning from a novelty into a commodity like water or power, one that now needs its own exchange and its own quality inspectors.

Original sources published: IT Home on-site report, Nov. 12, 2026, 10:52 Beijing time (UTC+8); PANews (citing STAR Market Daily), Nov. 12, 2026, 11:03 Beijing time; background: 36Kr, Nov. 5, 2026, 12:34 Beijing time