Baidu released ERNIE-5.5-Flash-Base on November 5 as a pretraining checkpoint for researchers and developers to adapt. The company publishes FP8 and BF16 weights. This report was published November 8.
A base checkpoint, not a finished assistant
The model card describes a text-only mixture-of-experts architecture with 280 billion total parameters and 14 billion activated per token. Baidu says training used up to 128K context and the released configuration supports a larger context for fine-tuning and inference. Context capacity is not a guarantee of usable speed or accuracy at that length.
Because the checkpoint is before the model's later training stages, it is a starting point for work such as instruction tuning, not the same thing as a ready-to-use chat product. Active parameters are not the same as the total weights that must be stored or offloaded; this is a very large artifact. We have not loaded or benchmarked it.
Read the actual license
The repository uses a custom Baidu license, not Apache or MIT. It contains territorial and commercial-use conditions, including exclusions for the EU, UK and South Korea in its defined territory. Check the full agreement for your location and application before use. Downloadable does not automatically mean unrestricted.
EXPLORE THE SOURCE
Projects in this story
Hugging Face models
These are the exact public repositories referenced in the coverage. Public weights may have restrictions: check each publisher’s terms.
Original reporting; claims attributed to their sources. No local tests implied.
