Baidu releases ERNIE 5.5 Flash Base with 280 billion parameters as an open-source foundation model

Baidu has published ERNIE-5.5-Flash-Base as an open foundation model, with weights available on Hugging Face in both FP8 and BF16 formats. The release is deliberately a base checkpoint: it ships without fine-tuning or post-training, leaving developers to decide what to build on top of it. Baidu says a post-trained version is planned to follow.
The model uses a mixture-of-experts architecture with 280 billion total parameters, of which 14 billion are active per input. It was pre-trained on 33.5 trillion tokens with a training context window of up to 128,000 tokens, and is designed to support contexts of up to one million tokens during fine-tuning and inference. The gap between the training context and the advertised one-million-token ceiling matters in practice: long-context ability at that scale depends on what developers do downstream, not just on the checkpoint as shipped.
Baidu reports that in benchmarks covering reasoning, code and long-context tasks, the model is competitive with similarly sized open foundation models, naming DeepSeek-V4-Flash-Base, Hy3-preview-Base and MiMo-V2.5-Base as comparisons. According to the company, all were tested with the same evaluation pipeline across Exam, STEM, Reasoning, Knowledge, Code, Multilingual and Long Context categories. These are vendor-reported figures, so the usual caution applies until independent evaluations appear.
The mixture-of-experts design is the key structural detail. Only a fraction of the 280 billion parameters fire for any given token, which keeps inference and fine-tuning costs closer to those of a much smaller dense model while retaining a large parameter pool. Serving a 280-billion-parameter model still requires substantial memory for the full weight set, so the FP8 release is likely to be the practical entry point for teams with limited GPU budgets, while BF16 suits those who want maximum numerical fidelity and have the hardware to match. This is my own reading of the trade-off, not something Baidu states.
Shipping a base model rather than an instruction-tuned one is a deliberate choice that shifts work onto the user. Base checkpoints are raw next-token predictors; they are useful for continued pre-training, domain adaptation and building custom post-training pipelines, but they will not behave like a chat assistant out of the box. Teams wanting a ready-to-use model will likely wait for the post-trained release Baidu has promised, while those with their own data and alignment recipes get a clean starting point.
Why it matters: developers weighing open-weight options now have another large Chinese MoE checkpoint to evaluate, at a size class where DeepSeek and Xiaomi's MiMo already compete. The base-only release suits teams with fine-tuning capacity, and the promised post-trained version will determine whether it becomes a drop-in alternative for product teams. The one-million-token context claim is the spec to watch, since it can only be validated once independent testing appears.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Baidu releases ERNIE 5.5 Flash Base: a 280-billion-parameter MoE model pre-trained on 33.5 trillion tokens, shipped deliberately without fine-tuning, so developers can take it wherever they want. The article Baidu releases ERNIE 5.5 Flash Base with 280 billion parameters as an open-source foundation model appeared first on The Decoder .