Baidu Releases ERNIE 5.5 Flash Preview: 280B Parameters, 14B Active, 1M Context

Chinese labs keep releasing large open-weight MoE models, and Baidu is adding a new variant to the lineup. The company has announced ERNIE 5.5 Flash Preview, the first model to undergo post-training on top of the previously released ERNIE 5.5 Flash Base. For now it is a preview, but the numbers in the announcement merit a closer look: the model is large in size and yet fast in operation.

What's in the announcement

Three parameters are publicly available so far. The first is the architecture: 280B parameters in total and 14B active. This is a Mixture-of-Experts design: the model contains many "experts," and only a small portion of them is switched on for each token. The second is the context window of 1 million tokens. The third is speed: more than 200 tokens per second per user on the Baidu Qianfan platform. And there is a promise: the weights will be opened by November 25.

The difference between "total" and "active" is important to grasp. Imagine a large library where, for each question, you open not all the cabinets but only the two or three you need. A lot of knowledge is stored, but little has to be read. As a result, the compute per token is roughly that of a 14B model, while the knowledge capacity is that of a noticeably larger one.

There is an important caveat: this does not save memory. To run the model yourself, you will still have to hold all 280B parameters in memory, even if quantized. The savings apply to speed and compute cost, not to storage size.

What "post-training" and "Preview" mean

A base model is the result of pre-training: it continues text but follows instructions poorly. Post-training (SFT, reinforcement learning and similar stages) turns it into an assistant that performs tasks, keeps the response format and calls tools. This stage is usually what decides whether a model is useful in a product. The word Preview means there is no final version yet and behavior may change.

One more detail: the post contains no benchmarks. There are no comparisons with Qwen, DeepSeek or GPT-class models, and no results on coding or agentic tasks. Therefore, claims about quality can neither be confirmed nor refuted for now. Only the architecture, context length and speed are known.

What developers and product managers should do

Until the weights are out, a sensible plan looks like this:

1. If you have access to Qianfan, run the model on your own tasks rather than on public benchmarks.
2. Test the long context in practice. The claimed 1M tokens does not guarantee that the model finds facts equally well at the beginning, middle and end of a document. Run a "needle in a haystack" test on real data.
3. Wait for the license. For commercial use it matters whether fine-tuning and resale are permitted, and whether there are restrictions by region or business size. The announcement says nothing about this.
4. Calculate the hardware. Even in 4-bit form, 280B is about 140 GB for the weights alone, meaning several server GPUs rather than a single workstation.

A speed of 200+ tokens per second matters for agents and interactive interfaces. Multi-step chains with dozens of model calls add up to seconds of waiting, and fast inference literally changes how a product feels. But note the wording "per user": this is the speed on Baidu's cloud stack under a specific load, and on your hardware the numbers will differ.

Bottom line

ERNIE 5.5 Flash Preview is not a breakthrough yet, as there are no independent measurements, but it is a strong signal: another large MoE model with a long context and a promise of open weights. For teams choosing an open model for agents or work with large documents, it is a shortlist candidate. The decision is best made after November 25, when the weights, the license and the first third-party tests appear.