Baidu releases ERNIE 5.5 Flash Preview, with model weights due November 25

Baidu released ERNIE 5.5 Flash Preview on November 12, 2026, giving developers API access to a model with a stated one-million-token context window. Hosted access is available now; downloadable weights are promised by November 25. The launch also brings two specialist models for expressive video creation and fast handling of complex requests.

Hosted access comes first

The preview is available through Baidu AI Cloud’s Qianfan platform, and Baidu says it is also free on OpenRouter. Those are the immediate routes for using the model. The company describes the release as open-weight, but its commitment to publish the model files is a separate, later step.

Qianfan pricing separates input, output and cached input, rather than charging one rate for all usage. Generated output carries the highest listed price. Cached input has its own substantially lower rate, making the distinction between those categories important when reading Baidu’s cost pitch. These are API prices, not terms for the planned weights release.

RMB 1Input

Baidu’s listed input rate on Qianfan.

RMB 4Output

Baidu’s listed output rate on Qianfan.

RMB 0.04Cached input

Baidu lists cached input separately from other input.

A large model built around inference efficiency

Baidu lists 280 billion total parameters, with 14 billion activated per token. The two figures distinguish the model’s full parameter pool from the portion activated for each token it processes. Its stated context window is one million tokens—the amount of material it can work with in context, rather than a measure of its parameter count.

The company positions Flash Preview for coding, 3D modeling, knowledge work and content creation. It says the model combines competitive results on tests including Terminal-Bench and AutomationBench with very fast decoding, or response generation. The announcement supplies neither numerical benchmark scores nor a measured decoding rate, leaving that performance comparison unquantified.

Behind the efficiency pitch is ERNIE Evolve, which Baidu describes as its proprietary recursive self-improvement technology. Baidu says it uses that technology to improve inference efficiency and reduce cost. That is the company’s explanation for the optimization; it is distinct from the published API rates, which specify what customers are charged.

Two specialist models take different routes

Alongside the foundation model, Baidu introduced ERNIE Wizshot Micro-Expression and ERNIE Light Deep Thinking. Both application models are available through Qianfan’s API, but their stated jobs differ: one focuses on expressive multimedia production, while the other is designed to answer complex requests and make tool calls at search-scale speed.

• Wizshot Micro-Expression processes text, images and speech together for expressive video creation. Baidu says it accepts text inputs of up to 4,000 Chinese characters and targets formats including podcasts and short animated dramas. The input limit is specifically stated in Chinese characters, not tokens.
• Light Deep Thinking was originally developed for multimedia search. It pairs a fast-thinking model that serves users with a slow-thinking model that continuously refines it in a closed loop. Baidu also says its Sparse Attention technology compresses historical memory during response generation to reduce inference costs without sacrificing performance.