Baidu announced on Tuesday, November 11, 2026, the launch of ERNIE 5.5 Flash Preview, the first post-trained version built on the ERNIE 5.5 Flash Base, whose open weights were released days earlier.
The model retains the sparse Mixture-of-Experts architecture with 280 billion total parameters, of which only 14 billion are activated per token, and offers a context of up to 1 million tokens. According to Baidu, generation exceeds 200 tokens per second, balancing capability and inference efficiency.
Focus on real productivity
The company positions the Preview for coding, 3D modeling, knowledge work and content creation tasks. In official demos, the model completed end-to-end projects inside agent frameworks, including creating a complete wuxia game in Unreal Engine 5.8 in an 11-hour coding session, parametric modeling of airports and airplanes in Blender, research report generation and animated video production.
Post-base training includes domain-specific data for game development (Unreal, Unity, Godot), professional tools like Blender and CAD, plus workspaces with hundreds of thousands of files for domains such as finance and law.
Availability and benchmarks
The API is already available on Baidu AI Cloud's Qianfan platform. Model weights are expected to be open-sourced under the Apache 2.0 license by November 25, 2026. The Preview can also be tested for free on OpenRouter.
In benchmarks published by Baidu, the model stands out on AutomationBench (61.9) and SpreadsheetBench V2 (36.4), outperforming some listed competitors, although it trails models like Claude Opus 5.5 on Terminal-Bench 4.0 (46.5 versus 63.6). Baidu notes that this is an intermediate post-training checkpoint and that training continues.
Sources: Official ERNIE announcement, Baidu post on X.
This article was produced with the assistance of artificial intelligence and editorially reviewed.
Por Grok
