ERNIE-4.5-VL-28B-A3B-Thinking

ERNIE-4.5-VL-28B-A3B-Thinking

Reads images and text and reasons about them, with about 3B of its 29.7B parameters active at a time. Covers chart reading, STEM problems from photos, visual grounding and zooming into images. Apache 2.0.

Made by baidu on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: This is Baidu's image and text reasoning model, an upgrade of ERNIE-4.5-VL-28B-A3B. It has 29.7B parameters in total and, per the card, activates only about 3B for each token. Baidu trained it further on a large set of visual reasoning data, then used reinforcement learning on tasks with checkable answers. It works in English and Chinese. The card says it matches top flagship models on many benchmarks, which are the authors' own claims.

Who it suits: It suits developers and researchers who build agents or tools that need to reason over pictures, charts and photographed problems.

Industry | Who uses it | How they use it | Example
Education | Teams building study tools | Solve STEM problems from photos | Read a photographed physics question and work out the answer
Business analysis | Analysts reading reports | Ask questions about charts and figures | Explain the trend in a sales chart
Industrial software | Developers of inspection tools | Use visual grounding to point at parts of an image | Locate a labeled part in a complex scene

You can skip it if you only work with text, since a text model needs far less memory than this image model.

Where to use it:

• Python scripts that use the transformers library
• Machines with enough memory for tens of GB of weights
• Windows, macOS and Linux with Python, to download the files

Use case | Who | What they get back
Visual reasoning | Developers | Step by step answers about what an image shows
STEM help | Students and tutors | Worked answers to problems in photos
Grounding | Tool builders | Locations in an image tied to an instruction

How to set it up: You need these first:

• Python on Windows, macOS or Linux
• Hardware that can hold about 59.3 GB of BF16 weights, or a smaller quantized copy
• The transformers library

1. Install or update the Hugging Face download client.
2. Download the model files to your computer.
3. Load the model in your own transformers code.
4. Read the full model card for the exact loading code.

pip install -U huggingface_hub
hf download baidu/ERNIE-4.5-VL-28B-A3B-Thinking

How to use it: You give the model an image and a question, and it replies in text. The card text available here shows no code, so follow the full card for how to load it and how to switch on its zooming feature.

Tool | What it does
Visual reasoning | Handles multi step reasoning, chart reading and cause and effect in images
Visual grounding | Ties an instruction to places in an image
Thinking with Images | Lets the model zoom in and out of an image to catch fine detail
Image zooming and image search | Tools the card pairs with that feature for detail and rare knowledge

Example: You photograph a physics problem and ask the model to solve it. The card names problems from photos as a STEM strength, so this is the kind of task it was tuned for.

For newcomers and experts: A reasoning model works through a problem before it answers. A good first try is giving it a chart and asking what trend it shows.

For readers who know the field, the card says the model went through mid-training on visual language reasoning data. Reinforcement learning then ran on verifiable tasks, with GSPO and IcePop used to keep mixture of experts training stable. It also used dynamic difficulty sampling.

Give the model an image and a question
It reasons over the picture
It may zoom into details
Optional tools add search or cropping
It writes the answer

Good to know:

• Claims: the comparison with flagship models is Baidu's own, and its benchmark picture is not in the text available here.
• Cut: the text available here ends in the list of key capabilities, before usage code.
• Size: 29.7B parameters, about 59.3 GB in BF16 or about 14.8 GB at 4-bit as an estimate.
• Active size: about 3B parameters are used per token, but all the weights must still be stored.
• Versions: Hugging Face lists 2 quantized and 2 fine tuned versions made by others.
• Language: Hugging Face lists English and Chinese.

Title: Weight size of the model
BF16: 59.3 GB
4-bit, estimated: 14.8 GB

• License: Apache 2.0.
• Cost to run: none named. You need hardware that fits the weights.
• Data it reaches: the images and text you give it. If you add an image search tool, that tool decides what leaves your machine.
• Not for: none named in the text available here.

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 29.7B
• Weights: about 59.3 GB in BF16; about 14.8 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 2,021
• Likes: 548
• License: apache-2.0
• Languages: en, zh
• Quantized versions: 2
• Fine-tunes: 2
• Tags: transformers, safetensors, ernie4_5_moe_vl, image-text-to-text, ernie4.5, conversational

Links

• Model: https://huggingface.co/baidu/ERNIE-4.5-VL-28B-A3B-Thinking
• Files: https://huggingface.co/baidu/ERNIE-4.5-VL-28B-A3B-Thinking/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download baidu/ERNIE-4.5-VL-28B-A3B-Thinking

Then follow the usage in its model card. Nothing in this kit runs by itself.