Baidu releases ERNIE 5.5 Flash Base: 280B parameters, open weights

Chinese labs continue to release large open models, and Baidu has rejoined their ranks with an unusual move. ERNIE-5.5-Flash-Base ships without instruction tuning: it is a raw base model that developers must turn into a working product themselves. The company promises to release a post-trained version later.

What is inside

The model uses a mixture-of-experts architecture: it has 280B parameters in total, but activates only 14B for each request. This matters in practice: quality is determined by the total size, while the compute cost per token is closer to that of a 14B model. However, you still need enough memory to hold all 280B parameters' worth of weights.

Pre-training covered 33.5 trillion tokens with a context of up to 128K tokens. The architecture is also designed to support contexts of up to 1 million tokens during fine-tuning and inference. That is Baidu's stated capability: users will have to extend the window themselves, and long-context quality should be validated on their own data.

The weights are available on Hugging Face in two formats: FP8 and BF16. FP8 has significantly lower memory requirements and is suitable for experiments on a smaller cluster, while BF16 is more convenient for fine-tuning.

What a "base" model is and how it differs from a chat model

A base model can continue text, but it has not been trained to follow instructions, reject harmful requests, or maintain a conversational format. Think of an intern with enormous knowledge whom nobody has taught how to respond to customers. Building an assistant requires supervised fine-tuning, usually followed by a preference-training or RL stage.

The base model does have one advantage: it contains none of someone else's decisions about style, filters, or behavior. A team seeking its own specialized model for legal documents, code in a rare language, or internal support gets a clean starting point.

Baidu compares the model with DeepSeek-V4-Flash-Base, Hy3-preview-Base, and MiMo-V2.5-Base using a single evaluation pipeline. The categories include exams, STEM, reasoning, knowledge, code, multilingual performance, and long context. According to the company, ERNIE performs on par with comparable open base models in reasoning, code, and long-context tasks. These are Baidu's own results; no independent measurements are available yet.

Limitations and what to do

First, without post-training, the model cannot simply be connected to a chat interface or an agent. You need data, a compute budget, and experience fine-tuning MoE models, which is more difficult than tuning a dense model.

Second, the license. Before commercial use, read the terms on the model page; the announcement does not specify them.

Third, infrastructure. Even in FP8, 280B parameters require several powerful GPUs, so this is not a local experiment for a small team.

The practical takeaway: if you have the resources and a specific use case, benchmark ERNIE-5.5-Flash-Base on your own workload alongside DeepSeek V4 Flash Base and MiMo V2.5 Base, then select the best foundation for fine-tuning. If you need a production-ready model, wait for Baidu's post-trained version: it has not been released yet, and no timeline has been announced.

Overall, this is another sign that open base models with hundreds of billions of parameters have become commonplace, while the competitive advantage is shifting away from the weights themselves and toward data and post-training quality.