OPEN LINK ↗
// 42m agoMODEL RELEASE
Baidu drops ERNIE 5.5 Flash Base
Baidu released an open-weight 280B-parameter MoE base model with 14B active parameters, trained on 33.5T tokens and supporting up to 1M-token contexts for fine-tuning and inference. It is a pretraining checkpoint without instruction tuning or reinforcement learning, designed for developers to adapt.
// ANALYSIS
ERNIE 5.5 Flash Base is an unusually developer-oriented release: Baidu is handing over a large, efficient foundation rather than a polished chatbot, but the restrictive license limits its global usefulness.
- –Sparse activation targets lower inference costs despite the model’s 280B total size
- –Repository-scale code data and long-context training make it promising for coding and agent research
- –The 1M-token inference context is valuable for large codebases and document-heavy workflows
- –Developers must add their own instruction tuning, alignment, and deployment stack
- –The custom license excludes the EU, UK, and South Korea, weakening its open-weight appeal
// TAGS
ernie-5.5-flash-basellmopen-weightsmoelong-contextfine-tuningtraining
DISCOVERED
42m ago
2026-11-05
PUBLISHED
3h ago
2026-11-05
RELEVANCE
9/ 10
AUTHOR
pretext