Baidu used its Baidu World 2026 conference on November 12 to formally launch Ciyuansuo, a platform it describes as "an exchange for large models" — a marketplace where enterprises and individuals can discover, buy and call tokens from multiple model providers under one roof. Founder Robin Li Yanhong framed the problem bluntly: in the first nine months of this year, new models shipped almost daily, yet reseller relays in between take no responsibility for whether the model behind an API is real, stable or fast. "Same model, some expensive, some cheap, sources all over the place," he said. "Users keep choosing and still do not dare use it."
The platform's answer is two mechanisms it calls verification and credit. Verification runs before a provider goes live: model fingerprint comparison to confirm the served weights match the advertised model, plus capability tests, performance benchmarks and availability checks. Credit is the ongoing layer — periodic sampling probes, full-link real-time call monitoring, and a standing rating that covers service quality, security compliance and commercial performance. The results accumulate into a public credit file for each supplier, so a provider that quietly swaps in a cheaper model or inflates context limits loses its grade in the open.
The trial numbers, company-reported and not independently audited, show the demand curve that motivated the build. In roughly one month of pilot operation, Ciyuansuo says it has onboarded more than 100 supply resources, passed 50 million calls per day, and moved over 700 billion tokens daily. Some 40-plus models are listed, spanning DeepSeek, Moonshot's Kimi, Zhipu GLM and MiniMax alongside Baidu's own ERNIE line, with both original model vendors and third-party inference providers among the sellers. A day after the formal launch, Baidu said more than 100 additional suppliers had applied to join and over 1,000 new customers had connected.
The market being formalized here already exists in gray form. As call volumes exploded, compute-rich intermediaries began reselling model access, and the layer grew fast precisely because even first-party APIs cannot always absorb peak demand. What that layer never had was accountability — no party stood behind the token when latency spiked or the model quietly degraded. A trust and ratings layer on top turns an opaque resale market into something closer to an exchange with listed counterparties.
Strategically, Ciyuansuo is also a distribution play. Alibaba's Bailian and ByteDance's Volcano Ark already aggregate multiple models for enterprise buyers, and a neutral marketplace gives Baidu a seat in routing traffic for models it does not own — while its ERNIE models sit on the same shelf as the rivals Chinese enterprises actually want to compare. The credit-ratings angle is the differentiator: nobody else is yet certifying that the DeepSeek endpoint you bought is actually running DeepSeek.
The open questions are enforcement and neutrality. Fingerprinting catches binary substitution but not subtler downgrades — quantized weights, trimmed system prompts, degraded sampling settings. And a platform that both sells its own models and rates competitors will face the usual exchange-conflict questions. Analysts quoted in Chinese coverage argue the deeper value is macro: idle mid-tier compute centers get a searchable, price-comparable outlet, which improves utilization across China's structurally mismatched compute supply. If the ratings regime holds up, token trading starts to look like a real commodity market; if it does not, Ciyuansuo becomes one more relay with better branding.
Comments (0)
Log in to join the discussion
Log InNo comments yet