Physix FrontierPhysix Frontier · News Briefing Card (Leiphone · Nov 5, 2026)
Baidu open-sources ERNIE 5.5 Flash Base, 280B-parameter MoE
KEY FACTS
- Baidu open-sourced its base model ERNIE 5.5 Flash Base on November 5.
- The model uses a MoE architecture with 280 billion total parameters and 14 billion activated per token.
- Pre-training used 33.5 trillion tokens, with a maximum context of 128K and inference support for 1 million tokens.
- The model achieves same-size SOTA on knowledge, reasoning, math, and code benchmarks.
- It outperforms comparison models including DeepSeek-V4-Flash-Base and Hy3-preview-Base.
KEY DATA
280 billionTotal parameters
14 billionActivated parameters
33.5 trillionPre-training tokens
128KPre-training context length
PHYSIX OBSERVATION
By open-sourcing a 280-billion-parameter base, Baidu shifts the competitive focus from stacking parameters to training and inference efficiency. For developers, there is one more strong fine-tunable base to choose from; for the industry, the engineering rollout of sparse attention and multi-route MoE could lower deployment costs in long-context scenarios, and its actual inference performance is worth watching.
Source: Leiphone report