The Pulse
Baidu previews ERNIE 5.5 Flash with 1M-token context
Baidu's ERNIE 5.5 Flash Preview is available through the Qianfan API with a 1M-token context. The company says the model is still in post-training and its weights will be released under Apache 2.0 soon.
Baidu has made ERNIE 5.5 Flash Preview available with a 1M-token context and 280 billion total parameters, of which 14 billion are active. The model's API is available on Qianfan, and its weights will be open-sourced under Apache 2.0 soon, the company said.
The November 11 release is the first post-trained model built on the earlier ERNIE 5.5 Flash Base. Baidu describes the preview as an intermediate checkpoint and says training is not finished, so some capabilities may still fall short.
Access and training
The company plans to monitor real-world use and feedback and make improvements in the official release. It says the preview is strong at coding, 3D modeling, knowledge work and content creation, and that it scores competitively across major benchmarks.
The release includes demonstrations across game development, design, knowledge work and video. Baidu says the model completed each example end to end inside an agent harness, with external tools and assets credited alongside the demonstrations.
Games and 3D design
In one game demonstration, Baidu says the model spent 11 hours merging assets from several generations of Unreal Engine into an Unreal Engine 5.8 project and writing a third-person game in C++. Because the model is text-only and cannot see the screen, it converted problems found during playtesting into measurable code tests, according to the company.
The reported problems included a hovering character, an inconsistent running direction and a misplaced camera. A growing test suite and playtest feedback guided iterations until the development goals were met, Baidu says.
Another demonstration used two agents running the preview with the Blender Python API: one built an airport and the other an aircraft. For the airport, the model found layout vector data on a mapping platform, downloaded supporting assets and built the terminal and airfield, according to the company.
The aircraft project assembled the fuselage and wings from cross-sections and used assets provided in advance for its interior. The model then matched the scale and cameras between the projects, arranged an animation and rendered it frame by frame, Baidu says.
Reports and video
For knowledge-work training, Baidu says it built 300,000 workspaces containing 10 million reference files in more than 40 file types. The environments ranged from personal workspaces with 10 to 100 files to enterprise knowledge bases with a thousand files or more.
Among the examples is a 20-page report on electricity costs associated with AI data centers in the United States. According to the company, the report used public research and regulatory materials and covered subjects including electricity demand forecasts, grid expansion, household bills and local fiscal benefits.
Baidu also describes a 116-second animated short written in Python. The film follows a pelican on a bicycle through different art styles, using more than 10,000 lines of code, with frames drawn and processed using external software libraries.
For three scenes, the model called an image-generation tool for static backgrounds and animated elements over them in code, the company says. The soundtrack used selected recordings edited to match scene changes.