Recently, Baidu released its latest foundation model, ERNIE 5.5 Flash Preview, with a total parameter count of 280 billion, activating 14 billion parameters per token, and supporting a context window of up to 1 million tokens. It stands out in coding, 3D modeling, knowledge work, and content creation, and has achieved competitive results on multiple benchmarks such as Terminal-Bench and AutomationBench.
Coding capability is the core highlight of this upgrade. The model covers a wide range of scenarios, from fixing and iterating on real-world codebases to building complete complex applications, and is stronger at planning, execution, and continuous iteration in long-horizon development tasks. Baidu conducted specialized training for game development and adapted the model to mainstream development tools, enabling it to complete game design, feature implementation, and scene construction based on user ideas.
In 3D modeling, Baidu strengthened the model's training on tools such as Blender and CAD, improving its performance in parametric modeling and spatial reasoning. With tool assistance, the model can infer an object's shape, proportions, and spatial relationships from text descriptions, generate accurate 3D models through procedural modeling, and integrate and reuse existing assets and external materials to build grand, detail-rich 3D scenes.
In video creation, the model can handle material gathering, compositing, and arrangement, adjust visuals and pacing, and output finished videos. Baidu has built more than 150,000 creation environments around scenarios such as science popularization, comic dramas, creative editing, and trending-topic adaptations, covering the full workflow of idea generation, scripts, storyboards, visual assets, and voiceover. It supports over a thousand creation skills and has produced about 5,300 minutes of finished videos in total.
The model is now available on Baidu AI Cloud's Qianfan platform, and developers can call it via API. Pricing is RMB 2 per million input tokens during peak hours and RMB 1 during off-peak hours, and RMB 8 per million output tokens during peak hours and RMB 4 during off-peak hours. It is currently available at a limited-time 50% discount, offering strong price competitiveness among models of similar parameter scale. The model weights will be released to the global open-source community on November 25.
Competition among large models is shifting from contests over general capabilities to deep scenario-based optimization, and open-source and low-price strategies are expected to further broaden developers' application space.
