Baidu open-sources ERNIE-4.5-VL-28B-A3B-Thinking

Background

Vision-language models take an image and a text prompt and answer in text. Many also write out intermediate reasoning steps before the final answer, which helps on tasks that need several rounds of reading and checking, such as pulling numbers out of a chart. Two terms are worth defining for this release. "Open-weight" means the trained weights are published, so others can download and run them. "Mixture of experts" (MoE) means the network holds many parameter blocks but routes each token through only a few of them, so the number of parameters in total and the amount of computation spent per token are different numbers.

What was released

In November 2025 Baidu released ERNIE-4.5-VL-28B-A3B-Thinking as an open-weight model ^[1][2]. The licence is Apache 2.0 ^[3]. The release came with training support rather than weights alone: Baidu added support for the model to ERNIEKit v1.5 ^[1][2], and the November 2025 news in the PaddlePaddle ERNIE repository lists ERNIEKit v1.5 as providing supervised fine-tuning and function-call training for it ^[1].

The name carries the specification. VL marks it as a vision-language model, 28B is the total parameter count, A3B means roughly 3 billion parameters are active per token, and Thinking refers to the reasoning mode. The Batch described it as a mixture-of-experts transformer with 28 billion total parameters and 3 billion active per token ^[2].

Thinking with images

The feature the model card singles out is that it can zoom into images and call image search while it reasons — what the card calls "Thinking with Images" ^[3]. The idea is that an image need not be read once and frozen into a fixed internal representation. If a detail is too small to read, the model can crop and enlarge that region and look again; if the picture shows something it cannot pin down from the pixels alone, it can issue an image search and reason over what comes back. Reading a dense chart, checking a label in the corner of a diagram, or identifying a building in a photograph are the kinds of step this loop is meant to support, with what it finds feeding back into the same chain of reasoning.

Training

The model card says the model was trained with multimodal reinforcement learning using GSPO and IcePop ^[3]. Reinforcement learning here means optimising the model against a reward signal rather than only imitating text: the model produces attempts, the attempts are scored, and training pushes probability toward the attempts that scored better.

Results

DeepLearning.AI reported a score of 87.1% on ChartQA, against 76.3% for Gemini 2.5 Pro and 78.2% for GPT-5 (high) ^[2]. ChartQA asks questions that can only be answered by reading values and trends out of a chart, so it tests the combination of perception and multi-step reasoning that this model is aimed at. A single benchmark is a narrow window on a model's abilities: it shows where this model stands on chart reading against those two systems, not a general ranking.

Running it

The Hugging Face model card states that single-card FastDeploy deployment needs at least 48GB of GPU memory ^[3]. That number is worth reading alongside the "3 billion active" figure. Sparse activation lowers the computation spent per token, but the full set of weights still has to be stored and served.

Why it matters

Three things travel together in this release. The weights are open and permissively licensed ^[3], so labs and companies can inspect, fine-tune and self-host instead of only calling a remote service. The training stack is published too, with ERNIEKit v1.5 covering supervised fine-tuning and function-call training ^[1], which is what a team needs in order to adapt the model to its own tools and data. And "thinking with images" treats the picture as something to revisit during reasoning rather than a single fixed encoding ^[3] — a design that suits charts, documents and screenshots, where the answer sits in a small region of a large image.