AI-NAV · Article
Baidu's Token Exchange: Why Ciyuansuo Marks the Model Industry's Next Phase

The most interesting product at Baidu's flagship event wasn't a model
Baidu World 2026, held in Beijing on 12 November, had all the usual ingredients of a Chinese tech keynote. A new foundation model, for one: ERNIE 5.5 Flash Preview, with 280 billion total parameters, 14 billion activated per token and a one-million-token context window. But the product that got the industry talking builds no models at all. It is a marketplace for them.
Robin Li, Baidu's founder, introduced Ciyuansuo — the name translates roughly as "Token Exchange" — with a single line: it is a stock exchange for large models. He leaned on the analogy himself. A stock exchange attracts good companies to list, so investors can find and back them. Ciyuansuo, he said, aims to gather good models and model providers, so that companies and even individuals can find, buy and use trustworthy model tokens.
The comparison is worth sitting with. An exchange produces nothing; it sets listing standards, enforces disclosure and lends credibility. With this launch, Baidu steps back from being purely a model vendor and takes up the position of market organiser. For a company whose AI reputation rests on ERNIE, that repositioning says more than any benchmark score on the slide deck.
What the exchange actually sells
Strip it down and Ciyuansuo (at cysuo.com) is a matching platform between model suppliers and model buyers. On the supply side sit model developers — DeepSeek, Moonshot's Kimi, Zhipu's GLM, MiniMax and, yes, Baidu's own ERNIE family — alongside third-party inference providers reselling capacity. On the demand side are enterprises and individual developers who compare, select, integrate and call models through one interface.
The homepage advertises three operating figures: over 200 billion tokens served daily, more than 60 million calls a day, and 70-plus model resources listed. A note on the numbers, because they differ from the keynote. Li's figures on 12 November described the one-month trial run: over 100 supply resources onboarded, daily calls past 50 million, daily token volume above 700 billion. The "70-plus" on the website counts listed model resources, while "100-plus" counts supply resources — the same model can be offered by several vendors. The call-volume gap likely reflects different measurement windows. These are separate metrics and should not be merged.
The commercial mechanics are deliberately plain. Register, verify your identity, top up, then browse a model market where the same model from different vendors is displayed side by side — input and output prices, first-token latency, throughput, seven-day success rate. Pick one, generate an API key, and you are calling within minutes. One integration covers every model on the shelf.
The price dispersion is real. As of mid-November, GLM-5.3-Flash started at 0.52 yuan per million input tokens with twelve vendors competing; DeepSeek-V4.1-Flash at 1.3 yuan with six; Kimi K3 at 13.68 yuan with four. Same model, multiple sellers, posted prices. That is what "exchange" looks like when it lands on a webpage.
Verification and credit scoring: the point is trust, not features
Aggregation alone would not be news. Alibaba's Bailian and ByteDance's Volcano Ark have offered multi-model platforms for a while, and OpenRouter has proven the model-routing model overseas. Ciyuansuo's pitch rests on two words: verification and credit.
Verification applies to the models. Before listing, a model goes through fingerprint comparison, capability testing, and performance and availability checks. After listing it faces periodic spot tests and full-chain real-time monitoring. The fingerprinting matters most: it is designed to stop a vendor quietly swapping the model behind the API — you think you are calling model A, but a cheaper model B is actually answering.
Credit applies to the vendors. Test and monitoring results feed into public credit ratings covering service stability, whether capabilities match claims, security compliance and commercial conduct.
Why make these the headline features? Li framed the problem bluntly on stage. In the first nine months of this year, more than one new model was released almost every day, and call volumes exploded. A crop of token resellers emerged to serve that demand — intermediaries that take no responsibility for model quality, service stability or response speed, and cannot even prove the model they sell is the model they claim. The same model at different prices from different sources, with uneven quality. Buyers browse, hesitate and walk away.
He is not alone in saying so. China's token-resale market has spent two years in a grey zone: inflated performance claims, swapped models and opaque billing were an open secret, and buyers had no practical way to audit any of it. Enterprise procurement fears the unverifiable far more than the expensive. Ciyuansuo turns verifiability into platform infrastructure. That — not the model count — is what separates it from every aggregator before it.
The market answered with real sign-ups
A platform is only as good as the two sides willing to show up. Figures Baidu released on 14 November, two days after launch: more than 100 new suppliers applied to join on day one, and over 1,000 new customers connected. By then the platform listed more than 40 models, covering the main Chinese text and video models plus ERNIE.
The supply-side enthusiasm is not hard to explain. Pan Helin, a member of the expert committee on information and communications economics under the Ministry of Industry and Information Technology, identified the underlying logic: China's core compute problem is structural mismatch — general-purpose compute in surplus, high-end intelligent compute scarce, and a cohort of small and mid-sized data centres holding capacity they cannot sell. A public marketplace gives that idle inference capacity a searchable, comparable outlet. For small providers sitting on unused servers, this is a sales channel that did not exist before.
Demand-side logic is just as direct. One API, side-by-side pricing and public success-rate data compress what used to be a two-week vendor evaluation into a single screen. Chen Xiaohua, executive dean of the education and research institute of the China Mobile Communications Association, read the launch as a marker that model commercialisation is shifting from single-vendor supply to platform-based market trading — a step towards transparent pricing and credible circulation in the token economy.
The industry just changed axes
Zoom out, and Ciyuansuo marks a migration in what China's model industry competes on.
For three years the axis was the model itself: parameter counts, benchmark scores, context length. But supply has swollen to the point where leadership windows are measured in weeks. When upstream supply commoditises, value moves downstream to distribution — whoever owns the trading gateway owns the steadiest position in the chain. Veteran telecom analyst Ma Jihua put it plainly: the platform helps users pick the right model, and in doing so, Baidu gains command over the gateway to those users. The flip side is a changed role for ERNIE — from an athlete that must win every model race to one listed asset among many in the exchange. That is both a hedge and a level-up.
Baidu did not invent this track and will not have it to itself. Alibaba's Bailian, Volcano Ark and the Alibaba-affiliated JellyToken all run multi-model aggregation; OpenRouter has validated demand abroad. Ciyuansuo's difference is that verification and credit scoring sit at the front door as public data rather than in the back office as risk control. Whether that difference becomes a moat depends on whether the ratings keep their credibility over time.
Three questions only time can answer
For an industry watcher, the unanswered questions matter more than the day-one sign-up numbers.
First, neutrality. ERNIE sits on the shelf, and the exchange is run by Baidu — the referee also fields a team. The site promises that all data is public, transparent and neutral, and for now ERNIE shows a single supplier against DeepSeek's six and GLM's twelve, a restraint that looks deliberate. But once the rankings carry real traffic, every weighting, every recommendation slot, every hint of favouritism toward the house model will be scrutinised.
Second, enforcement. When disputes arrive — a vendor's claimed version not matching what is served, or a fight over who owns an outage — can the platform genuinely arbitrate from the middle? A credit system only counts if the operator is willing to cut a top supplier.
Third, the price-war feedback loop. Transparent comparison will squeeze vendor margins. Good for buyers in the short run, but over time it could push smaller inference providers out and concentrate supply at the top — the classic exchange paradox, where better liquidity strengthens the strong.
Closing thought
Ciyuansuo signals that China's model industry has finished the first half — whose model is strongest — and started the second: whose market works best. What it builds is not technology but market infrastructure. Admission checks, live monitoring, credit ratings, transparent pricing: vocabulary borrowed from securities, e-commerce and payments, now imported into the token economy.
Three days in, the numbers — over a hundred supplier applications, over a thousand customers — suggest both sides had been waiting for this infrastructure. But an exchange is never judged by its opening-day fireworks. It is judged by whether the market still trusts its quotes and its ratings three years on. Baidu has made a good start. The rest of the exam is neutrality, and time.
Sources
- Baidu World 2026: Ciyuansuo and ERNIE 5.5 Flash Preview launched (China National Radio)
- Day one: 100+ new supplier applications, 1,000+ new customers (Huanqiu)
- Ciyuansuo official site
- Token Exchange product page (Baidu AI Cloud)
- Baidu's "exchange for large models" (China Business Journal)
- Ciyuansuo entry (Baidu Baike)
