Baidu open-sources ERNIE 5.5 Flash Base, a 280-billion-parameter model

Baidu has released ERNIE-5.5-Flash-Base, a 280-billion-parameter foundation model, as open source — and it ships deliberately without fine-tuning or post-training. That choice is the story. Most open-weight releases arrive ready to chat; this one arrives as raw material, and the developers who download it decide for themselves what it becomes.

Baidu says a post-trained version is supposed to follow. Until then, ERNIE 5.5 Flash Base is a foundation in the strictest sense: pre-trained, published, unfinished.

What exactly did Baidu release?

ERNIE 5.5 Flash Base uses a mixture-of-experts architecture with 280 billion total parameters, of which 14 billion are active per input. That gap between total and active parameters is the core engineering decision here. A dense model activates every parameter for every token it processes. A mixture-of-experts model routes each input to a small subset of its specialists, so Baidu gets the knowledge capacity of a very large model while paying the compute cost of a much smaller one.

Think of it as a hospital with 280 billion staff on the roster but only 14 billion on shift at any moment. The building can cover everything; each patient only sees a few doctors.

The model was pre-trained on 33.5 trillion tokens, with a training context of up to 128,000 tokens. During fine-tuning and inference, Baidu says the architecture is designed to support a context of up to one million tokens. The distinction matters: what a model saw during training and what it can handle when you actually run it are two different ceilings, and Baidu is promising the higher one for deployment rather than pre-training.

The weights are available on Hugging Face in two formats: FP8 and BF16. BF16 is the standard precision most of the ecosystem already speaks. FP8 is the smaller, cheaper format — it roughly halves memory and bandwidth requirements compared to FP16-class formats, which matters when you're hosting a 280-billion-parameter model. Offering both signals that Baidu expects serious deployments, not just hobbyist downloads.

How does it perform?

According to Baidu, the model competes with similarly sized open foundation models in benchmarks for reasoning, code, and long-context tasks. The company names two specific rivals: DeepSeek V4 Flash and MiMo V2.5.

That's a telling pair. DeepSeek has become the reference point for Chinese open-weight releases that punch above their price, and Baidu is explicitly placing itself in that league rather than claiming a leapfrog. The claim is parity with its size class, not dominance over it. Independent benchmark results aren't part of the announcement, so the performance picture rests on Baidu's own comparisons for now.

Why ship a model with no post-training?

The deliberate omission of fine-tuning and post-training is unusual enough to deserve a closer look. Post-training is the phase where a raw pre-trained model learns to behave like an assistant — following instructions, refusing harmful requests, adopting a persona, coding to spec. It's also where a company's own values, policies, and product decisions get baked in.

By stopping at the foundation stage, Baidu hands those decisions to whoever downloads the weights. A security firm can post-train it for code analysis. A research lab can align it differently. A company with its own compliance requirements can shape it to local rules.

There's a practical side too. Post-training costs money and time, and releasing the base model first gets the weights into developers' hands earlier. The announcement says a post-trained version will follow, so Baidu is splitting the release in two: raw capability now, polished behavior later.

The takeaway: Baidu is betting that a segment of developers wants the substrate, not the product — and that the ecosystem around ERNIE 5.5 will grow faster if builders can shape it from the ground up.

What does this mean for the open-source AI race?

The release is a signal that Baidu's AI development is still alive — the company remains a serious participant in the open-weights competition, even as much of the burden shifts to the developers doing their own post-training.

The competitive frame Baidu chose — DeepSeek V4 Flash, MiMo V2.5 — places this release inside an increasingly crowded middleweight division of open models: large enough for demanding work, small enough in active parameters to run affordably. Mixture-of-experts designs with low active-parameter ratios have become the standard play in that division, because they let labs advertise frontier-class total capacity while keeping inference costs in range of commercial deployments.

The 33.5-trillion-token pre-training run also tells you where the cost sits in modern model development. The expensive, compute-heavy phase is done and published. What Baidu has outsourced is the cheaper but more opinionated phase — deciding what the model should be.

For developers, the practical questions are straightforward: whether the million-token context holds up in real deployments, whether FP8 weights run cleanly on their hardware, and whether the claimed parity with DeepSeek V4 Flash and MiMo V2.5 survives third-party testing.

The next data point is already promised: when the post-trained version of ERNIE 5.5 lands, we'll see whether Baidu's open-first, finish-later strategy turns downloads into a real ecosystem — or just into weights sitting unused on Hugging Face.