ERNIE-4.5-VL-28B-A3B-PT

ERNIE-4.5-VL-28B-A3B-PT is a multimodal model from the ERNIE 4.5 family, built on a heterogeneous Mixture-of-Experts (MoE) architecture. It has 28 billion total parameters, with only 3 billion activated per inference pass, ensuring high computational efficiency. A key innovation lies in its modality-specific expert groups: separate experts handle textual and visual inputs, while shared experts and self-attention layers enable effective cross-modal interaction. The model features an adaptive vision encoder that processes images at arbitrary resolutions without distorting their aspect ratio, preserving the original proportions. For video, it employs an adaptive frame sampling strategy with temporal timestamps rendered directly onto frames, enabling precise temporal understanding. It supports a context window of up to 131,072 tokens, allowing it to handle lengthy documents and extended video sequences.

The model offers two operational modes—thinking and non-thinking—making it versatile across diverse tasks. The thinking mode enhances reasoning for complex visual challenges (e.g., STEM, mathematics, puzzles), while non-thinking mode enables rapid processing of simple, routine requests. Compared to the flagship ERNIE-4.5-VL-424B-A47B, this compact 28B-A3B variant exhibits only marginal performance degradation while drastically reducing computational requirements.

The model’s multimodal capabilities enable a wide range of practical applications: its strong performance on OCRBench (885) and DocVQA (94.1) demonstrates effectiveness in processing scanned documents, invoices, and forms; high scores on ChartQA (82.2) and TableVQA (70.0) make it suitable for analyzing charts and tables in financial and scientific data; its video understanding capabilities (MVBench 72.0, VideoMME 74.4, LongVideoBench 62.1) are valuable for security and surveillance systems; and its precise object counting (CountBench 87.6) and visual perception (RealWorldQA 69.2) can be leveraged in retail for inventory management and visual search. Released under the permissive Apache 2.0 license, the model can be freely used in commercial projects without restrictions.

Announce Date: 28.06.2025
Parameters: 29B
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-PT 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.