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

ERNIE-4.5-VL-28B-A3B-Thinking is based on an innovative heterogeneous Mixture-of-Experts (MoE) architecture. In this architecture, textual and visual inputs are routed to separate sets of experts, each specialized for the characteristics of its respective modality (modality-isolated routing). Integration is achieved through shared self-attention layers for all modalities and a group of shared experts. The visual experts have one-third fewer parameters than the textual experts, enabling efficient processing of visual information while reducing computational costs by approximately 66% for visual tokens. This architecture prevents errors in processing different data modalities while ensuring high efficiency during the semantic sequence formation stage when they are combined. The architecture includes an adaptive Vision Encoder based on ViT, which processes images with arbitrary resolution while preserving their original aspect ratios, and also supports video through an adaptive frame sampling strategy with temporal markers. The model supports a context window of 131,072 tokens, allowing it to process long documents and extended video clips.

The key distinction of ERNIE-4.5-VL-28B-A3B-Thinking from the base version and other models in the lineup is its specialized additional training for multimodal reasoning tasks, achieved through an extensive mid-training phase on high-quality visual-linguistic data. The model utilizes advanced multimodal reinforcement learning techniques (GSPO and IcePop) on verifiable tasks, including visual STEM problems and visual puzzles. It supports the unique "Thinking with Images" feature—an ability to "think" in a human-like manner by zooming in on images and capturing their details for subsequent analysis, additionally, it can utilize tools for image search within the problem-solving process (this requires integrating an external function).

ERNIE-4.5-VL-28B-A3B-Thinking is excellently suited for a wide range of tasks requiring deep understanding of multimodal data. It is particularly effective in recognizing and interpreting documents – from financial reports and scientific articles to engineering drawings and tables. Thanks to its thinking mode, which provides step-by-step reasoning, the model can be used within educational applications. Its video understanding capabilities make it useful for video surveillance systems, sports analytics, and media content cataloging. The model is built and trained on the PaddlePaddle framework; however, there are versions of the weights that are supported by all popular modern frameworks for inference and fine-tuning. Furthermore, it is distributed under the open-source Apache 2.0 license, making it freely available for commercial use.

Announce Date: 07.11.2025
Parameters: 30B
Experts: 130
Activated at inference: 3B
Context: 132K
Layers: 28
Attention Type: Full Attention
Developer: Baidu, Inc.
Transformers Version: 4.57.6
License: Apache 2.0

Use our pre-built public endpoints for free to test inference and explore ERNIE-4.5-VL-28B-A3B-Thinking capabilities. You can obtain an API access token on the token management page after registration and verification.

An endpoint is a server with pre-installed software that automatically downloads neural network weights and launches an OpenAI-compatible API to handle incoming requests.

A private endpoint is created within the user's project, belongs exclusively to the user, and is billed per server configuration in accordance with the infrastructure pricing.

We recommend creating private endpoints when you need to:

• achieve maximum endpoint performance,
• access the full context window for processing large requests and conversations,
• ensure maximum security by processing data in a dedicated, isolated environment.

Contact our dedicated neural networks support team at support@immers.cloud or send your request to the sales department at sale@immers.cloud.