ERNIE 5.5 Flash Base

ERNIE 5.5 Flash Base is a language model Baidu released on 5 November 2026 as a base checkpoint: the weights at the end of pre-training, before any mid-training, instruction tuning or reinforcement learning. Baidu presents it as a common starting point for domain adaptation, instruction tuning and reinforcement-learning research, so that developers do not have to pay for large-scale pre-training themselves. It is not a chat model and is not offered through a priced API on the release page.

## A sparse model built to be cheap to train further

It is a mixture of experts with 280 billion parameters in total, 14 billion of them active for each token, in 43 Transformer layers. Baidu says it was pre-trained from scratch on 33.5 trillion tokens with a training context of up to 128,000 tokens, and that it supports a one-million-token context for fine-tuning and inference. To keep long contexts affordable, only one layer in seven keeps token-level access to the whole context through multi-head latent attention; the others compress every 128 tokens into a single cached entry and attend over those plus a short window of recent tokens. Baidu says the pre-training corpus emphasises abilities that are hard to add later, such as repository-scale code and long-context understanding.

## Where to get it

The weights are on Hugging Face under the baidu organisation, for the transformers library, with an FP8 copy, in English and Chinese. The licence is Baidu's own ERNIE 5.5 Flash Base Model License Agreement v1.0 rather than a standard open-source licence, so its terms should be read before commercial use. Baidu's evaluation results on the release page are its own and are not stored as scores here.

## Key facts

**Developer:** Baidu **Released:** 5 November 2026 **Size:** 280 billion parameters, 14 billion active (mixture of experts) **Training data:** 33.5 trillion tokens (company figure) **Context:** 128,000 tokens in training; 1 million for fine-tuning and inference **Licence:** ERNIE 5.5 Flash Base Model License Agreement v1.0 **Form:** base checkpoint, no post-training

Not yet independently verified: Not yet independently verified. The architecture, training data size and evaluation results are Baidu's own; we found no independent test or dated report, and we have not read the licence in full. We will update this when it can be confirmed, and remove this note.