Baidu Open-Sources ERNIE 5.5 Flash Base, Reporting Same-Size Benchmark Gains

What happened

Baidu open-sourced ERNIE 5.5 Flash Base on November 5, reporting state-of-the-art results among same-size models across knowledge, reasoning, math, coding, multilingual, and long-context benchmarks. The model uses a mixture-of-experts architecture with 280 billion total parameters and 14 billion active per token; its pretraining used 33.5 trillion tokens. The article describes attention, expert routing, and training changes intended to improve resource efficiency.

Read next

• More changes in today’s briefingOpenAI Releases Images 2.6 With SVG Export, GIFs, and Faster Generation

Reported by 2 sources

The story so far

Baidu has open-sourced ERNIE 5.5 Flash Base, a 280-billion-parameter mixture-of-experts base model that activates 14 billion parameters per token. The model is available on Hugging Face. Baidu says the model was pretrained on 33.5 trillion tokens, supports 128K context during pretraining and up to 1 million tokens during fine-tuning and inference, and achieved top results among similarly sized models across several benchmarks.