On November 11, Baidu’s latest basic model Wenxin 5.5 Flash Preview was officially released. Wenxin 5.5 Flash Preview has a total parameter volume of 280 billion, each Token activates 14 billion parameters, supports a context of up to 1 million Tokens, and has outstanding performance in programming, 3D modeling, knowledge work, content creation, etc. Currently, Wenxin 5.5 Flash Preview has achieved competitive scores in multiple benchmark tests such as Terminal-Bench and AutomationBench, and maintained the ultimate decoding speed.

Programming ability is a highlight of Wenxin’s new model. Wenxin 5.5 Flash Preview covers a wide range of coding scenarios, from repair iterations in real code bases to complete construction of complex applications. Especially in long-term development tasks, the model shows stronger planning, execution and continuous iteration capabilities, which also greatly improves the game development effect of Wenxin 5.5 Flash Preview. The model has undergone special training for game development scenarios and is adapted to mainstream development tools. It completes game design, function production and scene construction based on users' creativity and needs, allowing users to more easily turn game ideas into experiential works.

At the same time, Wenxin’s new model is particularly outstanding in parametric modeling and spatial reasoning. By strengthening the model's special training in using tools such as Blender and CAD, Wenxin 5.5 Flash Preview's 3D modeling and design capabilities have been significantly improved. With the help of tools, Wenxin 5.5 Flash Preview can accurately understand and infer the shape, proportion and spatial relationship of objects from text descriptions, and convert them into accurate 3D models through procedural modeling. At the same time, it can flexibly integrate and reuse existing assets, integrate external materials into existing workflows, and create 3D scenes with grand scenes and rich details.

This capability also extends to professional knowledge work and content creation. Wenxin 5.5 Flash Preview can complete multi-step work according to user needs and deliver highly professional forms, documents or presentations. For fields such as finance and law, it has built 300,000 workspaces and 10 million reference documents, covering more than 40 file types. The training environment covers personal workspaces from 10 to 100 files, as well as an enterprise knowledge base composed of thousands of files.

For the video creation needs of content creators, Wenxin 5.5 Flash Preview can collect, synthesize and arrange materials, adjust the picture and rhythm, and output the film. It has built more than 150,000 creative environments for scenes such as science popularization, comics, creative editing, and hot-spot interpretations, covering the complete process of creative generation, scripts, storyboards, visual materials, and dubbing. It is adapted to thousands of creative skills, and the total length of the film is about 5,300 minutes.

It is understood that in terms of underlying technology, Wenxin 5.5 Flash Preview improves model capabilities while reducing computing and memory overhead through a number of technological innovations such as a new hybrid sparse attention architecture, multi-router MoE, and ECHO multi-Token prediction. It supports millions of Token long contexts and achieves a generation speed of more than 200 tokens per second. In addition, Wenxin 5.5 Flash Preview introduces self-developed RSI technology for the first time, and ERNIE Evolve participates in model inference and optimization, exploring new ways for AI to participate in model research and optimization.

Currently, Wenxin 5.5 Flash Preview has been launched on Baidu Smart Cloud Qianfan platform, and developers can call the model through API. The API pricing is 2 yuan per million tokens when input is busy and 1 yuan when idle, and 8 yuan when output is busy and 4 yuan when idle. It is a 50% discount for a limited time. It has strong price competitiveness in the same parameter scale model. In addition, model weights will be officially released to the global open source community on November 25.

edit

Resend Email Modify Email