ERNIE 5.5 Flash Preview

Arena.ai's weekly roundup shows Anthropic's Claude Fable 5.2 taking first place in Agent Arena with a 30.5% net improvement across 7.8k real-world sessions, while Google DeepMind's Gemini 4 Hydrogen becomes the fastest top-10 model at a typical 7.3 seconds per step. Meta's smallest model Muse Ember debuted in Code Arena WebDev, Thinky Machines' Inkling 2 Small leads US open models, and Baidu's ERNIE 5.5 Flash Preview nears second among open models with a 1M-token context.

Baidu unveiled ERNIE 5.5 Flash Preview, a 280B-parameter MoE model activating only 14B per pass, with a 1M-token context and over 200 tokens/sec output at as little as ¥1 per million input tokens off-peak. It claims top scores on Terminal-Bench 4.0 and AutomationBench, demoed an Unreal Engine 5 wuxia game scene built in 11 hours, and will open-source its weights globally on November 25.

Baidu has unveiled ERNIE 5.5 Flash Preview ahead of its World Conference, featuring 280B total parameters with only 14B activated per inference, million-token context, and over 200 tokens/sec generation, with an open-source release promised. The model centers on coding capability, spanning game development, 3D modeling and video orchestration that require sustained tool use, and Baidu claims it rivals flagship models like Claude Opus 5.5 and GPT-6.1 Sol on benchmarks such as Terminal-Bench and AutomationBench.

Baidu has released ERNIE 5.5 Flash Preview, the first post-trained variant of its Base model, featuring 280B total parameters, 14B activated parameters and a 1M-token context window, targeting coding, 3D modeling and content creation. The API is live on the Qianfan platform, with weights promised to be open-sourced by November 25; Baidu describes it as an intermediate post-training checkpoint still under rapid iteration.