THE ESSENTIALS IN 30 SECONDS

  • ERNIE 5.5 Flash Base has 280B total parameters, 14B active per token, and was pretrained from scratch on 33.5T tokens, according to its model card.
  • It is a base checkpoint with no mid- or post-training; Baidu says a post-trained version will follow but has given no date.
  • The custom licence is free for non-commercial use and for commercial use up to $50 million in related annual revenue, and does not cover the EU, the UK or South Korea.

Baidu has published the weights of ERNIE 5.5 Flash Base, a large text-only language model built on a sparse mixture-of-experts (MoE) design. The model card on Hugging Face lists 280 billion total parameters, of which about 14 billion are active for each token, and says the model was pretrained from scratch on 33.5 trillion tokens. Baidu dated the release and its licence November 5.

A raw checkpoint, not a chatbot

The release is unusual for a major lab: it is the checkpoint taken at the end of pretraining, before any mid-training or post-training. In practice, it completes text rather than following instructions, and is meant for teams that want to run their own instruction tuning, domain adaptation or reinforcement-learning experiments. Baidu says a post-trained version will follow, without a date.

According to the model card, the network has 43 layers, 512 routed experts with ten selected per token plus one shared expert, and a vocabulary of about 201,000 tokens. It was trained with a 128K-token context, while the released configuration enables a 1M-token context for fine-tuning and inference. Weights come in FP8 and in a separate BF16 repository, and Baidu provides a plugin for the vLLM serving engine on GitHub. OrcaRouter estimates the FP8 weights alone need about 280 GB of GPU memory.

Baidu compares the model with other recent open base models of similar size, including DeepSeek-V4-Flash-Base. In OrcaRouter's reading of those tables, ERNIE leads on 24 of 30 rows. These scores are reported by Baidu and have not yet been reproduced independently.

A licence with conditions

Although widely described as open source, the model ships under a custom "ERNIE 5.5 Flash Base Model License Agreement v1.0" rather than a standard open-source licence. Non-commercial uses such as research, education and evaluation are free. Commercial use is free while annual revenue from related products stays at or below US$50 million, on condition of a one-time notice to Baidu and a visible "Powered by ERNIE" attribution. Above that level, separate written authorisation is required.

The licence territory covers the whole world except the European Union, the United Kingdom and South Korea, so organisations there are not authorised to use or redistribute it. That is a notable change from June 2025, when Baidu released its ERNIE 4.5 family under Apache 2.0.