Baidu releases 280-billion-parameter ERNIE 5.5 Flash Base model

Summary

Baidu has open-sourced the ERNIE 5.5 Flash Base model, featuring a 280-billion parameter MoE architecture. Utilizing 14 billion active parameters per token, the model was pre-trained on 33.5 trillion tokens and offers a 1-million-token context window. The company aims to establish a cost-effective foundation for the upcoming full release, 'ERNIE 5.5 Flash.'

Why it matters

Releasing a 280-billion parameter MoE base model with a 1-million-token context window as an open-weight asset intensifies the open-model race with DeepSeek, Tencent, and others. This move provides developers with a large-scale starting point without ownership restrictions. The goal of minimizing downstream fine-tuning and inference costs signals that long-context capability is becoming a default feature rather than an add-on. Since a full version with an instruction-tuned variant is promised to arrive soon, this base version should be read as the first act rather than the final product.

Highlights

• 280-billion total / 14-billion active parameter MoE architecture pre-trained on 33.5 trillion tokens
• 1-million-token context window embedded in the base model for inference and fine-tuning
• Self-reported benchmarks claim it outperforms similarly scaled models like DeepSeek-V4-Flash-Base, Hy3-preview-Base, and MiMo-V2.5-Base

Critical take

All performance claims rely on Baidu's own benchmark table and a single tweet describing the model as 'SOTA for its size.' The article lacks independent evaluation, ablation studies for the 1-million-token context window, or details on licensing and computational costs. Until third-party reproductions are available, the 'superior' framing remains essentially vendor marketing.