Baidu releases ERNIE-5.5-Flash-Base for further training

Baidu announced ERNIE-5.5-Flash-Base as a base model for developers to train further.[1] The company says releasing it is intended to spare developers the cost of large-scale pretraining.[1] The weights are the checkpoint at the end of pretraining, before mid-training or post-training.[1] That distinction matters for teams assessing how much additional work may be needed: Baidu describes the model as a starting point, and says developers can use it for domain adaptation, instruction tuning and reinforcement-learning research.[1] The Decoder reports that the weights are available on Hugging Face in FP8 and BF16 formats.[2]

ERNIE-5.5-Flash-Base is a sparse mixture-of-experts language model with 280 billion total parameters and 14 billion activated per token.[1] Baidu says it pretrained the model from scratch on 33.5 trillion tokens.[1] The company also says efficiency was a design goal from the outset.[1] These specifications give teams concrete points to include in an evaluation, but the parameter counts alone do not establish performance or the compute requirements of a particular deployment. Developers should assess the model against their own workloads and infrastructure needs.

Baidu says the model was trained with a context length of up to 128,000 tokens.[1] Separately, it says the model supports a one-million-token context for fine-tuning and inference.[1] Those figures describe different stages and uses, so teams should not treat the one-million-token figure as the context length used for pretraining. The Decoder reports that the released weights come in FP8 and BF16 formats.[2] For engineering teams, the stated context support and available formats are practical items to test against intended fine-tuning and inference workflows.

Baidu says ERNIE-5.5-Flash-Base competes with similarly sized open foundation models such as DeepSeek V4 Flash and MiMo V2.5 on reasoning, code and long-context benchmarks.[2] That is the company’s characterization of benchmark results; the evidence provided here does not include scores or independent comparisons. The Decoder reports that a post-trained version is supposed to follow.[2] For now, the released checkpoint is deliberately without fine-tuning or post-training.[2] Teams considering it for a product should therefore evaluate it as a base model and account for the additional training work their use case requires.