The problem Token Exchange is actually solving
Model reselling has a trust problem that is easy to state. A buyer pays for a named model and has no simple way to confirm that the endpoint answering is that model, at that quality, at that speed. Baidu's answer is an exchange with an inspector attached. Before a resource is listed, Baidu says it compares the model's fingerprint and tests capability, performance and availability. After listing, the checks continue: regular spot checks, plus monitoring of calls from end to end, feeding a credit rating for each provider.
That design is the substantive part of this launch. A one-time listing check would not stay true for long in a market moving as fast as this one: Science and Technology Daily's account of Robin Li's Baidu World keynote, republished through ChainCatcher, reports his statement that more than one new model was released almost every day in the first nine months of this year. Continuous monitoring is at least aimed at the right failure mode.
Read the launch numbers as one company's count
The scale figures arrived without a second source. More than 100 resources, a daily average of more than 50 million calls, over 700 billion tokens a day, about a month in: every number in that sentence comes from Baidu's own release, and no auditor, customer or independent measurement was cited with it.
Dividing the rounded headline figures — 700 billion by 50 million — gives about 14,000 tokens per call. Treat that as desk arithmetic on Baidu's rounded numbers, not a reported average: both totals are stated as lower bounds, so the true ratio is not established. What the release does not do is explain its token accounting or divide the volume into paying enterprise traffic, free-tier testing, and Baidu's own internal calls. The split is not published.
Secondary coverage already shows how loosely these numbers travel. The KuCoin/ChainCatcher roundup of the same announcement put "500 million daily calls" in its headline, while its own body text reported daily requests exceeding 50 million. A tenfold error in a headline is a small thing until it gets quoted as the figure. Baidu reports a daily average of more than 50 million calls; that is the number to use.
"Neutral" is the word to test
Baidu describes Token Exchange as "a neutral, transparent, and trusted marketplace." Transparent can be checked over time: either the pricing and performance data stay public and current, or they do not. Neutral is harder. Baidu operates the exchange, runs the Qianfan cloud, sets the verification tests, assigns the credit ratings, and sells ERNIE models that compete with the third-party services being graded.
None of that makes the ratings wrong, and the release does not establish whether its tests favor Baidu's own services. It makes the grades Baidu's, produced under a conflict the announcement does not address. Baidu's release publishes neither the fingerprinting method nor the rating formula, so for now a provider's grade is a Baidu judgment a buyer is asked to trust — which is the same arrangement, one level up, that the exchange was built to replace.
The model that came with it
ERNIE 5.5 Flash Preview, released a day earlier, is, in Baidu's words, the first post-trained model built on the ERNIE 5.5 Flash Base that Baidu shipped in early November. It keeps the base model's shape, 280 billion total parameters with 14 billion active per token, with a 1-million-token context window, and is priced on Qianfan at RMB 1 per million input tokens and RMB 4 per million output, with cached input at RMB 0.04. Baidu's model page calls it "an intermediate checkpoint from post-training" and cautions that because training is not finished, "some capabilities may still fall short."
The license is the dated promise to watch. The Base model shipped on November 5 under Baidu's own custom license. For the Preview, Baidu's English model page promises the weights will be open-sourced under Apache 2.0, and the November 12 release dates that to by November 25, 2026. If the weights land on time and the license file shipped with them matches the promise, the Preview's terms will be materially more open than its base's. Until then it is an API with a price, sold in the same market where Baidu has just appointed itself referee.
Sources
- [1] Baidu, "Baidu Unveils Latest ERNIE Models Amid Broader Full-Stack AI Lineup, and Debuts Token Exchange to Power Intelligent Organizations" (PR Newswire, November 12, 2026)Read source
- [2] Baidu, ERNIE 5.5 Flash Preview release page (English, November 11, 2026)Read source
- [3] Unite.AI, "Baidu Launches Token Exchange for Verified AI Model Services" (November 12, 2026)Read source
- [4] KuCoin News / ChainCatcher roundup, "Baidu Launches Large Model Trading Platform 'Ciyuansuo' with Over 100 Resources and 500 Million Daily Calls" (November 12, 2026; headline figure conflicts with its own text, which reports 50 million — cited only for that discrepancy, and for the Science and Technology Daily account of the keynote)Read source