Baidu opens an AI model marketplace with continuous quality control
Baidu introduced Token Exchange, a marketplace to compare and purchase AI model services. The tool comes with new components for agents, cloud, and chips, with scaling already visible.
Baidu launched Token Exchange, a marketplace to compare and purchase AI model services, at Baidu World 2026 on November 12, 2026. The Chinese group pairs it with updates on its models, agents, cloud, and chips, following a very simple logic: making AI more purchasable, measurable, and verifiable. According to Baidu, the platform already handles on average more than 50 million calls and over 700 billion tokens per day, a volume that speaks to the ambition of the project.
A marketplace that does more than just display prices
Token Exchange allows both companies and individuals to compare prices and performance, then access services from multiple model providers. Baidu emphasizes especially the verification mechanism: comparison of model fingerprints, capacity, performance, and availability tests. Providers also receive credit scores based on regular spot checks and real-time end-to-end call monitoring. More than 100 model resources have already been integrated, placing the platform beyond a simple directory.
The choice is interesting because it addresses a problem that has become concrete for AI buyers: a model can look good on paper but be less reliable in production. Baidu therefore tries to establish a continuous trust layer, rather than a one-time validation at deployment. It’s less glamorous than a spectacular demo, but much more useful when signing a contract.
Broader, faster models designed for agents
At the same time, Baidu unveiled ERNIE 5.5 Flash Preview, presented as the first post-trained model built on ERNIE 5.5 Flash Base. This model features 280 billion parameters in total, 14 billion activated per token, and a context window of one million tokens. Baidu positions it for code, 3D modeling, knowledge work, and content creation.
The most telling demonstration remains that of a development agent: on an 11-hour Unreal Engine task, the text model wrote a third-person game in C++ then transformed test problems into measurable scenarios, unable to see the screen. Baidu says it bets on a compromise between capacity and inference efficiency, with an API accessible via Qianfan at 1 RMB per million input tokens and 4 RMB per million output tokens, or 0.04 RMB for cached inputs. The weights are to be released under Apache 2.0 by November 25, 2026, although the company warns this is an intermediate checkpoint still subject to improvement.
Agents, cloud, and chips: Baidu wants to connect the entire chain
The other announcements show the same layering logic. ERNIE Wizshot Micro-Expression combines text, images, and voice for expressive video creation. ERNIE Light Deep Thinking handles complex search queries and tool calls by combining a fast user-oriented model and a slower model that continuously refines the response. Baidu also says its Sparse Attention technology compresses historical memory during decoding to reduce inference costs without degrading performance.
On the agents side, Robin Li, co-founder and CEO of Baidu, advocated an organization focused on tasks rather than fixed hierarchies. His three principles are summed up in three verbs: break down work boundaries, transform tacit knowledge into shared assets, and create measurable process loops. Internal figures presented by Baidu serve as a commercial argument: daily active agents rose from 5,000 to 37,000 in one year, while weekly completed tasks climbed from 84,000 to over 1 million.
In the same logic, DuMate becomes the unified entry point to Baidu’s specialized agents, with a Harness engine and an Agent API open to developers. The team claims to have iterated more than 160 times in six months and written 90% of its code with the agent. KOOKO, from GenFlow 3.0, mixes public information, professional databases, and user content to produce research reports, presentations, and websites. Famou 3.0, for its part, targets industrial operations with an agent platform and engineering services deployed on-site; Baidu says it has clients in 25 of the 41 Chinese industrial categories.
Hardware support remains at the core of the plan
The infrastructure aspect completes the picture. Baidu AI Cloud launched Tianchi 1024-card Supernode, an all-optical interconnection system designed for high-speed data transfer. Kunlunxin, the group’s AI chip branch, says it has shipped its M100 inference accelerator, while the M300 is expected in 2027 with announced gains in computing power, memory, and interconnection.
Taken together, these launches outline a fairly coherent chain: chips, cloud, models, agents, then distribution marketplace. Token Exchange pushes Baidu beyond its own products by organizing comparison between providers, which may interest companies wanting to arbitrate between several offers without playing the benchmark lottery. One more delicate question remains: will these credit scores and continuous monitoring be enough to establish the platform as a trusted channel? [to be verified]
Key points
- Token Exchange was presented at Baidu World 2026.
- More than 100 model resources are already integrated.
- 50 million daily calls, according to Baidu after one month.
- Daily active agents have risen to 37,000.
- Robin Li advocates a task-centered organization.
By the numbers
- 700 billion tokens per day — average volume announced by Baidu after one month.
- 280 billion parameters — ERNIE 5.5 Flash Preview, presented on November 12, 2026.
- 1 million tokens — context window of ERNIE 5.5 Flash Preview.
- 90% of the code — share of DuMate team’s code generated with the agent, according to Baidu.
- 25 industrial categories — Famou 3.0 clients in China, out of 41 categories in total.