Baidu announced ERNIE 5.5 Flash Preview and Token Exchange at Baidu World 2026 in Beijing on 12 November. The foundation model is available through the company’s Qianfan AI cloud platform, while the marketplace is intended to connect model service providers with customers — giving organisations a way to buy access to models as well as use Baidu’s own.
Key points
- Baidu says ERNIE 5.5 Flash Preview has 280 billion parameters, activates 14 billion per token and supports a 1 million-token context window.
- Its API costs RMB 1 per million input tokens, RMB 4 per million output tokens and RMB 0.04 per million cached input tokens, according to Baidu.
- Baidu has launched Token Exchange to connect model providers and customers; its reported usage figures come from the company’s announcement.
- Baidu plans to release the Preview model’s weights by 25 November.
ERNIE 5.5 Flash Preview activates 14 billion parameters
Baidu describes ERNIE 5.5 Flash Preview as a model with 280 billion parameters in total, of which 14 billion are activated for each token it processes. Think of the larger figure as the stock of machinery available and the smaller one as the machinery switched on for a particular piece of work: the design draws on only part of the model at each step, although the full set of weights remains part of the model. Baidu says this arrangement is intended to make inference more efficient.
The company says the model can take up to 1 million tokens of context in a request. That is the space available for material supplied alongside an instruction, such as the text a model is asked to examine. Baidu also says it used its ERNIE Evolve technology to improve inference efficiency and reduce cost, and reports competitive results on Terminal-Bench and AutomationBench alongside fast decoding.
Reading a long document could involve putting more of its text into one request if it fitted within that context window. The amount submitted would still count towards input-token charges, which Baidu has set separately from its price for generated output.
On Qianfan, Baidu prices the API at RMB 1 per million input tokens and RMB 4 per million output tokens, with previously cached input priced at RMB 0.04 per million tokens. Those rates give companies building applications a stated price for sending material to the model and a different one for receiving its response. Baidu says the model is also available free on OpenRouter.
The Preview designation carries a qualification. Baidu characterised it as an intermediate stage of post-training, said some capabilities may be incomplete and expects further improvements in an official version, TokenPost reported. That matters for work in which a model’s behaviour must be assessed before it is incorporated into an application: the version available through the API is still undergoing that process, according to Baidu’s account.
Token Exchange connects providers and customers
Token Exchange extends the announcement beyond Baidu’s own models. The company describes it as a marketplace where model service providers can reach customers seeking access to their services. Baidu calls the marketplace neutral and says it uses model fingerprint verification and provider credit ratings to support trust between the parties.
After one month, Token Exchange had brought on more than 100 model resources, averaged more than 50 million calls a day and processed over 700 billion tokens daily, StreetInsider reported. Those are Baidu’s figures for activity on its marketplace, rather than measurements of how any individual model performs.
For companies building on models, the marketplace and Qianfan serve different purposes within the same announcement. Qianfan provides the API for ERNIE 5.5 Flash Preview at Baidu’s published rates. Token Exchange, as Baidu describes it, brings together services from model providers and prospective customers, making the choice of a provider part of the company’s wider offering.
Baidu adds Wizshot, Light Deep Thinking and agents
Baidu also introduced two models for particular applications, both available through Qianfan APIs. ERNIE Wizshot Micro-Expression takes text, images and speech as inputs for video creation, according to the company, and can accept long-form text of up to 4,000 Chinese characters. ERNIE Light Deep Thinking is designed for complex requests and calls to other tools.
For Light Deep Thinking, Baidu describes a pairing: a fast-thinking model handles requests while a slow-thinking model continuously refines it. The company says its Baidu Sparse Attention technology compresses earlier material held in memory during decoding to lower inference costs. These application models sit alongside the foundation model rather than replacing its general-purpose API.
Baidu announced upgrades to DuMate, its general-purpose agent; KOOKO, a workspace for documents and creative work; Baidu Yijing, a digital human agent; and Famou 3.0, a self-evolving agent. The additions follow other recent releases: AI Affairs covered Baidu’s Unlimited OCR 2 for parsing long documents in one pass and Apollo World for training robotaxis on rare road hazards.
Baidu plans to release the weights for ERNIE 5.5 Flash Preview by 25 November.
