Baidu releases open weights for ERNIE 5.5 Flash Base
ERNIE 5.5 Flash Base has 280 billion total parameters with 14 billion activated per token, and supports a 1-million-token context for fine-tuning and inference.
AT A GLANCE
Key points
- The model has 280 billion total parameters, with 14 billion activated per token.
- The checkpoint contains pretraining only, with no mid-training or post-training.
- FP8 and BF16 weights are available, with a 1-million-token context supported for fine-tuning and inference.
What is changing?
Baidu has released the weights for ERNIE 5.5 Flash Base. The sparse Mixture-of-Experts base model has 280 billion total parameters, with 14 billion activated for each token. The company says it was pretrained from scratch on 33.5 trillion tokens with a training context of up to 128K tokens.
The released checkpoint represents the end of pretraining and has not undergone mid-training or post-training. It is therefore positioned as a starting point for domain adaptation, instruction tuning, and reinforcement-learning research rather than as a ready-made chat assistant. Fine-tuning and inference support a 1-million-token context.
FP8 and BF16 safetensors weights are available on Hugging Face under the ERNIE 5.5 Flash Base Model License Agreement v1.0. Baidu also provides serving instructions for SGLang and vLLM. The benchmark results in the announcement are Baidu’s own evaluations, not independent verification.
Explore the original sources
Source date: 5 November 2026.
Published by AI Radar: 5 Nov 2026, 14:08 (Istanbul, UTC+3).
FOLLOW THE TOPIC
Topics and tags
Share your discovery.
Send this story to a friend.
Reader comments
What do you think about this development? Share your experience or question. Comments are published after editorial approval.
Your name/nickname and comment become public if approved. Read: Privacy notice.
Loading comments…