Subscribe
Explainers
9 min read
axc
axc

What Is ERNIE? Baidu's AI Models Explained

A plain-English guide to ERNIE: the name, ERNIE 4.5, 5.0, 5.1, the new 5.5 Flash Base weights, the open image model, and three ways to use Baidu's models as of November 2026.

Olivetti-style cover reading WHAT IS ERNIE with the subtitle Baidu AI models, explained, beside four bars labeled 5.5, 5.1, 4.5 and IMG in teal, red and mustard on cream paper
Olivetti-style cover reading WHAT IS ERNIE with the subtitle Baidu AI models, explained, beside four bars labeled 5.5, 5.1, 4.5 and IMG in teal, red and mustard on cream paper

ERNIE is the name of Baidu's family of AI models. As of November 2026, ERNIE 5.1, released on May 9, 2026, is the newest ERNIE model that Baidu offers through a chat site and an API. On November 5, 2026, Baidu also released the weights of ERNIE 5.5 Flash Base, a base model that is not tuned for chat. Baidu says a post-trained ERNIE 5.5 Flash will follow and gives no date. Baidu also publishes open-weight models under the ERNIE name, so you can download some of them. This guide explains what the family contains, how the names work, and three ways to use it.

Everything below is dated "as of November 2026" and links to Baidu's own pages. Baidu changes model lists often, so check the linked page before you build on a detail.

The short version

  • ERNIE is a family of models, not one product. It covers text chat, image understanding, image generation and more.
  • The newest ERNIE model with a chat site and an API is ERNIE 5.1, from May 9, 2026. Baidu's post links a chat site and a playground, not downloadable weights.
  • Most other open-weight models are older or smaller. They include the ERNIE 4.5 family from June 30, 2025, and an image model called ERNIE-Image.
  • ERNIE 5.5 Flash Base is the newest open-weight release, from November 5, 2026. It is a 280 billion parameter base model under a custom license, not Apache 2.0. Baidu says a post-trained version will follow.
  • Developers can reach ERNIE 5.1 through Baidu's cloud platform, Qianfan, with the model ID ernie-5.1.

Where the name comes from

ERNIE started as a research paper. In 2019, researchers at Baidu published ERNIE: Enhanced Representation through Knowledge Integration. The name is short for "Enhanced Representation through kNowledge IntEgration."

That first ERNIE was a language model of the same size as BERT-base, with 12 layers. Its idea was a new way to train. Instead of hiding single words and asking the model to guess them, it hid whole phrases and named entities, such as a person or a book title. The paper reports better results than BERT on five Chinese language tasks. It is a long way from today's chat models, but it explains why Baidu still uses the name.

A short timeline

These are the main steps, with the date from Baidu's own pages.

  • 2019: the original ERNIE paper, described above.
  • June 30, 2025: Baidu announced the open-source release of ERNIE 4.5, a family of 10 models, under the Apache 2.0 license.
  • February 2026: Baidu published the ERNIE 5.0 technical report, and its ERNIE blog post is dated February 6, 2026.
  • April 15, 2026: Baidu introduced ERNIE-Image, an open image model.
  • May 9, 2026: Baidu released ERNIE 5.1.
  • November 5, 2026: Baidu released the weights of ERNIE 5.5 Flash Base, a base model, under a custom license.
Timeline of ERNIE from the 2019 paper to the ERNIE 5.5 Flash Base open weights on November 5, 2026, with the open-weight releases marked

ERNIE 4.5: the open family

ERNIE 4.5 is the part of the family you can download. Baidu's announcement says it has 10 variants. Baidu describes most of them as "mixture of experts" models, plus one small dense model. Baidu's ERNIE 5.0 post explains the idea for its own model: it holds a very large pool of parameters but activates only a small share for each token, which Baidu says keeps computational cost comparable to much smaller dense models.

Baidu says the largest ERNIE 4.5 model has 424 billion total parameters, and the smallest is a dense model with 0.3 billion. The text model ERNIE-4.5-300B-A47B has 300 billion total parameters and 47 billion active for each token, according to its Hugging Face model card. The "A47B" in the name means 47 billion active.

Baidu says all of the ERNIE 4.5 models are available under Apache 2.0. Read the Apache 2.0 license text before you use a model commercially.

Baidu later added "Thinking" versions. Hugging Face lists ERNIE-4.5-21B-A3B-Thinking and ERNIE-4.5-VL-28B-A3B-Thinking, both under Apache 2.0. The first is a text model. The second reads images as well as text. Baidu's Qianfan documentation says that a deep-thinking model outputs a chain of reasoning before its final answer.

Baidu's launch post also names two tools. ERNIEKit is for fine-tuning, and FastDeploy is for running the models. The post says FastDeploy has an API compatible with the vLLM and OpenAI formats. Both tools rest on PaddlePaddle, Baidu's own deep learning framework. Baidu also offers PyTorch-compatible weights, which carry "-PT" in their Hugging Face names, for people who work outside PaddlePaddle.

ERNIE 5.0: one model for text, images, video and audio

ERNIE 5.0 is a different kind of model. Baidu's report says it was trained from scratch to handle text, images, video and audio in one system. Many multimodal models start as a text model and add image or audio parts later. Baidu calls that approach "late fusion" and says ERNIE 5.0 avoids it.

Baidu's blog post says ERNIE 5.0 has 2.4 trillion parameters. The report says fewer than 3% of them are active for each token. So the model is very large on paper but uses a small slice at a time.

The report also describes "elastic training." The idea is to train one large model so that smaller versions can be cut out of it later without a separate training run. Baidu reports that a version using 53.7% of the active parameters and 35.8% of the total parameters kept near-full performance in its tests. These are Baidu's own results.

ERNIE 5.1: the newest model for chat and API use

ERNIE 5.1 builds on ERNIE 5.0. In its release post, Baidu says it cut the total parameters to about one third and the active parameters to about one half of ERNIE 5.0's. Baidu also says it reached "leading foundational performance at its model scale" while using only about 6% of the pre-training cost of comparable models.

Those are Baidu's claims, and the post does not name the models it compared against for the cost figure. Treat them as a vendor statement.

On agent tasks, Baidu says ERNIE 5.1 surpasses DeepSeek-V4-Pro on two tests, τ³-bench and SpreadsheetBench-Verified, which it calls agent evaluation tasks. Baidu also says its math score on AIME26 with tool use was 99.6. Independent testers have not confirmed these figures in the sources used here, so read them as marketing until others repeat them.

Baidu's post also says ERNIE 5.1 would be rolled out progressively on more than ten creative production agent platforms. It names ISEKAI ZERO, Mulan AI, Diting Huanliu and Storymaster as examples. Baidu's Qianfan model list files ERNIE 5.1 under a heading that translates as "ERNIE series flagship models", and it lists no ERNIE 5.5 model.

ERNIE 5.5 Flash Base: open weights, base model only

On November 5, 2026, Baidu released the weights of ERNIE-5.5-Flash-Base on Hugging Face and described them in an ERNIE blog release post. It is a sparse mixture-of-experts text model with 280 billion total parameters and 14 billion active for each token. The weights are the checkpoint at the end of pre-training, before any mid-training or post-training. So it is a starting point for further training, not a chat model. Baidu says the post-trained ERNIE 5.5 Flash will follow and gives no date.

The license is custom: the ERNIE 5.5 Flash Base Model License Agreement v1.0, not Apache 2.0. Non-commercial use is free. Commercial use has conditions, including a revenue threshold. The license also sets a territory that excludes the European Union, the United Kingdom and South Korea. Read the license text on Hugging Face before you use the model.

ERNIE-Image: a small open image model

ERNIE-Image is a text-to-image model. Baidu's announcement says it has 8 billion parameters in its main image network, plus an optional 3 billion parameter "prompt enhancer" that expands a short prompt into a longer one. Baidu says it runs on consumer-grade hardware with 24 GB of video memory.

Baidu highlights text inside images, such as posters, in both Chinese and English. It publishes two versions: ERNIE-Image, and ERNIE-Image-Turbo, which Baidu says needs only 8 steps. Hugging Face lists both as Apache 2.0. Baidu's benchmark tables are its own, so treat its claim of leading performance among open-weight models as Baidu's.

Other Baidu models you may meet

Not every Baidu model carries the ERNIE name. Baidu's Hugging Face page also lists document-reading models, such as Unlimited-OCR-2, which Hugging Face lists under the MIT license. The llms.blog explainer on Unlimited OCR 2 covers that model in detail. Baidu's ERNIE blog also covers PaddleOCR-VL-1.5, a 0.9 billion parameter model for parsing documents. Baidu's Hugging Face organization page lists these and other models.

Three ways to use ERNIE

Three ways to use ERNIE: chat, the Qianfan API, or download open weights, with the model each one reaches

1. Chat with it

Baidu's ERNIE 5.1 post tells readers to visit the ERNIE website to chat with ERNIE 5.1. It also says Baidu AI Studio launched an ERNIE 5.1 playground. Sites and sign-in rules can change, so follow the links in Baidu's post for the current route.

2. Call it from software

Baidu's cloud platform, Qianfan, lists ERNIE 5.1 in its model list, which shows an update date of November 3, 2026. The list is in Chinese. It gives the model ID ernie-5.1, a context length of 128k tokens, a maximum input of 119k tokens, and an output range of 1 to 65,536 tokens.

The same Qianfan list also offers models from other labs, including DeepSeek, Qwen, GLM and Kimi. So Qianfan is a general model marketplace as well as a home for ERNIE. For prices and sign-up rules, use Baidu's own pages, because they change.

3. Run an open model yourself

You can download the open ERNIE models from Hugging Face. The size you pick decides the hardware. The model card for the 300 billion parameter text model shows a vLLM command for 16 GPUs with 80 GB each. A small model is far easier: Baidu's launch post shows a local example with the 0.3 billion parameter ERNIE-4.5-0.3B on FastDeploy. Match the model to your machine before you download. ERNIE-5.5-Flash-Base has 280 billion total parameters, so check its model card for serving options and hardware first. It is a base model, so it is not tuned for chat.

How ERNIE compares with other Chinese model families

China has several large model families. China's AI Labs Explained covers them side by side. For other Chinese companies, see the guides to ByteDance's Doubao chatbot and the news on Tencent's Hy4 model.

One pattern is visible in Baidu's own pages. The open-weight models named above are the ERNIE 4.5 family, ERNIE-Image and, since November 5, 2026, ERNIE-5.5-Flash-Base. Baidu's ERNIE 5.1 post links a chat site and a playground. Neither the ERNIE 5.0 post nor the ERNIE 5.1 post links downloadable weights, and Baidu's Hugging Face page lists no ERNIE 5.0 or 5.1 model. That can change, so check the model card for each new release.

What to check before you rely on it

  • Version names. Baidu's blog shows ERNIE-5.1-Preview on April 30, 2026 and ERNIE 5.1 on May 9, 2026. Use the exact model ID from the platform you call.
  • Vendor claims. Most benchmark figures here come from Baidu. Look for independent tests before you choose a model for work.
  • Licenses. Apache 2.0 and MIT cover most of the open models named here. ERNIE-5.5-Flash-Base uses a custom license with a revenue threshold and a territory limit, so read it before commercial use. None of these licenses cover ERNIE 5.1, which is not offered as a download in the sources used for this guide.
  • Language. Some Baidu documentation exists only in Chinese. A browser translation tool helps, but check numbers on the original page.
  • Dates. This guide reflects Baidu's pages as of November 2026. Baidu says a post-trained ERNIE 5.5 Flash will follow, so the model list may change soon.

Sources

ExplainersChinaOpen Weights

Written by

axc
axc

Edited by the llms.blog desk.

More in China

All China →

Popular this week