Baidu opens ERNIE application models for search tasks and expressive video
Chinese technology company Baidu has made two application-focused ERNIE models available through its Qianfan cloud platform. One combines fast responses with feedback from a slower model for complex requests and tool use. The other processes text, images and speech for expressive video. Enterprises and partners can access both through programming interfaces following their joint announcement at Baidu World.
Artificial Intelligence··Night
Two application models reach Qianfan
Chinese technology company Baidu made two application-focused ERNIE models available through its Qianfan cloud platform at Baidu World on November 12.[1], [2]
Enterprises and partners can access the models through programming interfaces. One is intended for complex search requests and tool use, while the other generates expressive video from combined media inputs. The video application model is named ERNIE Wizshot Micro-Expression in Baidu’s announcement.[1], [2]
Fast responses receive slower feedback
The search model combines a fast-thinking system serving live requests with a slower model supplying feedback. Baidu Sparse Attention compresses previously cached computation during generation. Baidu developed the search-focused model for Search, extending its tasks into answers, tool calls and multistep work. Thinking and non-thinking modes are offered, with coding, logical reasoning and text processing among the stated uses. The company’s descriptions of speed and cost advantages remain its own claims.[1]
Video combines speech and expression
Wizshot processes text, images and speech together to turn emotional cues into facial expressions and body movement. It accepts scripts of up to 4000 characters. Passage-level dubbing voices a complete script in one pass, rather than assembling shorter recordings, with the intention of maintaining tone and pacing. Qianfan’s product description says the model can animate a still portrait into a speaking video lasting 4–15 seconds. Both application models formed part of the same launch.[1]
