Baidu has put the weights of ERNIE-5.5-Flash-Base on Hugging Face, The Decoder reported on 5 November. The model has 280 billion parameters, but Baidu says only 14 billion work on each token. Baidu says it learned from 33.5 trillion tokens of text and read up to 128,000 tokens at a time. It says tuning and use can stretch to 1 million tokens. The weights come in FP8 and BF16, The Decoder said.
This is the raw model, saved before any tuning, and Baidu says it left that step to developers. The Decoder says a tuned version should follow, but neither source gives a date. Baidu says most layers keep one cache entry (a memory of earlier text) per 128 tokens of input. It says it planned the design to hold down memory costs for further training and long inputs. The sources do not describe the license terms.
Baidu supplies the general ability and leaves the model's behaviour to the buyer, so tuning, testing and safety work moves in-house. Only 14 billion parameters are active per token, yet the sources do not say what hardware holds all 280 billion. Budget owners can ask who will do the post-training, what hardware fits the model, and what the license allows.