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

What it is

ERNIE-4.5-VL-28B-A3B-Thinking is an open-weight multimodal reasoning model released by Baidu in November 2025^[1]. The name is a description of the architecture: 28 billion total parameters arranged as a mixture of experts (MoE), of which only 3 billion are activated for each token the model processes^[3]. It takes text, images and video as input, up to 128,000 tokens, and returns text^[2]. The "Thinking" suffix marks it as a model that works through an intermediate reasoning step before giving its answer.

Where it comes from

It is built on the ERNIE-4.5-VL-28B-A3B mixture-of-experts architecture^[3], and Baidu published the weights under the Apache License 2.0^[3]. Alongside the model, Baidu's ERNIEKit training toolkit moved to version 1.5 in the same month, adding supervised fine-tuning and function-call training for this model^[1].

Thinking with images

The capability Baidu puts at the centre of the model card is what it calls "Thinking with Images": in the middle of a reasoning pass the model can zoom into part of an image and later zoom back out, and it can call tools such as image search^[3]. This matters for inputs where the answer sits in a small region — the axis labels of a dense chart, a line of text in a screenshot — because zooming gives the model a second, higher-resolution look at the same pixels instead of forcing it to decide from one downsampled view.

The model also does visual grounding, locating objects named in a prompt, and video understanding with event localization, meaning it can indicate when in a clip something happens rather than only describing the clip as a whole^[3].

How it was trained

Baidu trained it with multimodal reinforcement learning, using two techniques, GSPO and IcePop, described in the model card as stabilising reinforcement learning for the MoE architecture^[3]. Keeping reinforcement learning stable is awkward on a mixture of experts, where each token touches only a slice of the network, so the training signal has to land in the right places; that is the problem the two techniques are reported to address.

Benchmarks

Baidu reported ChartQA 87.1%, against 76.3% for Gemini 2.5 Pro and 78.2% for GPT-5 (high)^[2]. On OCRBench it scored 858, against 810 for GPT-5 (high) and 866 for Gemini 2.5 Pro^[2]. Read together the two results are mixed: ahead on chart question answering, behind Gemini 2.5 Pro on text recognition in images.

Running it and paying for it

With the weights open, the model can be self-hosted. Single-card deployment with FastDeploy requires at least 48GB of GPU memory^[3]. The recommended vLLM serving setup uses an ernie45 reasoning parser and an ernie45 tool-call parser, with automatic tool choice enabled^[3]. For fine-tuning, ERNIEKit supports supervised fine-tuning, LoRA and DPO^[3]. Anyone who would rather not host it can call it through Baidu's Qianfan platform, where DeepLearning.AI reported pricing of $0.14 per million input tokens and $0.56 per million output tokens^[2].

Where it sits

A 28-billion-parameter model that activates 3 billion per token, and runs on a single card with 48GB of memory, is in the range where individuals and small teams can serve it themselves^[3]. The release followed a familiar pattern: weights, training toolkit, deployment recipes and a hosted endpoint arriving together, so the same model can be fine-tuned, self-hosted or called as a paid service. What is worth recording for readers following open multimodal models is that Baidu published the weights and the benchmark numbers behind its claims together, including the comparison in which the model trails Gemini 2.5 Pro^[2].