The ERNIE 5.5 Flash Base weights are now public. Baidu released the 280-billion-parameter model on Hugging Face on November 5, 2026, and you can download it today. But it is a raw base model, so it won’t chat with you out of the box.
That makes it a tool for developers and researchers, not a new chatbot for everyday users. Here’s what you actually get, what’s missing, and who should care.
Here are the facts that matter, per The Decoder and Baidu’s Hugging Face model card:
- Released November 5, 2026, with weights in FP8 and BF16 formats.
- 280B total parameters, with about 14B active for each token.
- Trained on 33.5 trillion tokens, text only.
- Custom “ERNIE 5.5 Flash Base Model License Agreement v1.0,” not a standard open license.
- A post-trained (chat-ready) version is expected later, but no date is given.
If you only read one line, make it the license one. “Open” here means downloadable, and the terms still need a read before you build a product on it.
What the ERNIE 5.5 Flash Base model actually is
ERNIE 5.5 Flash Base is a mixture-of-experts model. Only a small slice of its 280B parameters, roughly 14B, runs for each token, which keeps it cheaper to run than a dense model of the same size.
“Base” means it is the pre-training checkpoint. It has had no instruction tuning or post-training, so it predicts the next word rather than following requests. Baidu says it is meant as a starting point for fine-tuning and research.
The Hugging Face page also notes you reach it through a completions endpoint, not a chat interface. If you’ve used chatbots only, that’s the main difference you’ll feel.
Why release a base model at all?
Teams that build their own assistants usually want to shape the model themselves. A base model gives them a clean starting point, with no one else’s tone or safety rules baked in.
They can fine-tune it on their own data, such as support tickets, legal documents or code. Researchers also like base checkpoints because they show what pre-training alone produced.
The trade-off is effort. You need data, GPUs and time to turn it into something people can talk to. That’s why the chat-ready version matters more to most readers.
Long context and size at a glance
Baidu trained the model on contexts up to 128K tokens. The released config enables 1M tokens for fine-tuning and inference.
| Spec | ERNIE 5.5 Flash Base |
|---|---|
| Total parameters | 280B (Hugging Face metadata lists 286B, counting embeddings and output head) |
| Active per token | About 14B |
| Training data | 33.5 trillion tokens |
| Training context | Up to 128K tokens |
| Fine-tuning / inference context | Up to 1M tokens (enabled in the released config) |
| Formats | FP8 and BF16 safetensors |
| Serving options | SGLang and vLLM |
The 1M figure is a design target, not a tested promise. Baidu trained at 128K, so real long-document quality is something you should test yourself.
How good is it? Only Baidu’s word so far
Baidu says the model competes with similar open base models, including DeepSeek-V4-Flash-Base, Hy3-preview-Base and MiMo-V2.5-Base, across seven categories. Those include reasoning, code and long context.
The Decoder didn’t list exact scores, and the claims haven’t been independently checked. Treat them as marketing until outside benchmarks show up.
If you follow open models closely, this fits a busy stretch. We covered Inkling 2 Small, which is also a large open-weights release that arrived with caveats.
Can you run it at home?
Probably not on a normal PC. Baidu didn’t publish hardware requirements, but its serving examples use 4-way parallelism, which points to a multi-GPU server.
The vLLM example also needs a plugin from Baidu’s GitHub repository. And the model card says no inference provider currently hosts it, so you can’t just call it through an API yet.
If you want something smaller for a single GPU, our guide to running Qwen3.8-27B locally is a more realistic starting point.
Who should try this open model, and who should wait
Pick the group that sounds like you:
| If you are… | Our take |
|---|---|
| An ML team with GPU servers | Worth testing as a fine-tuning base, after you read the license. |
| A researcher | Good fit. A raw pre-training checkpoint is what research needs. |
| A hobbyist with one GPU | Wait for smaller or quantized versions. |
| Someone who just wants a chatbot | Skip it. Wait for the post-trained release. |
Frequently Asked Questions
Is ERNIE 5.5 Flash Base free to download?
Yes, the weights are on Hugging Face. But they come under Baidu’s own license, so check its terms before commercial use.
Can I chat with ERNIE 5.5 Flash Base?
Not in a useful way yet. It is a base model with no post-training. A chat-ready version is expected later, per The Decoder.
What does “280B total, 14B active” mean?
It’s a mixture-of-experts design. The model holds 280 billion parameters, but only about 14 billion are used for any one token, which saves compute.
Does it really support 1 million tokens?
The released config enables 1M tokens, but training used up to 128K. Test it on your own documents before you rely on it.
My verdict: this is a solid release for people who fine-tune models, and a non-event for everyone else. Wait for the post-trained version and independent benchmarks before you judge it, and read the license before you ship anything with it.
Leave a Reply