ERINE-4.5

Free

Open-source multimodal MoE model family with state-of-the-art performance

FreeFree tier

Inputs: text, imageOutputs: text

Type

Open Source

Company

Baidu

Visit Website GitHub (40 stars)

About ERINE-4.5

ERNIE 4.5 is a family of large-scale multimodal models developed by Baidu, comprising 10 distinct variants including MoE (Mixture-of-Experts) models with 47B and 3B active parameters (largest 424B total) and a 0.3B dense model. It features a novel heterogeneous modality structure for joint training on text and visual modalities, enabling parameter sharing across modalities while maintaining dedicated parameters per modality. Built on the PaddlePaddle deep learning framework, ERNIE 4.5 achieves 47% Model FLOPs Utilization (MFU) in pre-training and supports high-performance inference across diverse hardware platforms. The models are open-sourced under Apache 2.0, accompanied by industrial-grade development toolkits for resource-efficient training and inference.

Key Features

Multimodal Heterogeneous MoE Pre-Training jointly trained on text and visual modalities

Heterogeneous modality structure with parameter sharing and dedicated parameters per modality

MoE models with 47B and 3B active parameters (largest 424B total parameters) and a 0.3B dense model

Built on PaddlePaddle deep learning framework for efficient training and inference

Scaling-efficient infrastructure: heterogeneous hybrid parallelism, FP8 mixed-precision training, fine-grained recomputation

Modality-Specific Post-Training using SFT, DPO, and Unified Preference Optimization (UPO)

High-performance inference with multi-expert parallel collaboration and 4-bit/2-bit lossless quantization

PD disaggregation with dynamic role switching for effective resource utilization

State-of-the-art performance across text and multimodal benchmarks (instruction following, world knowledge, visual understanding, multimodal reasoning)

Open source development toolkits with industrial-grade capabilities and multi-hardware compatibility

Pros & Cons

Pros

Open source under Apache 2.0 with public access to model weights
State-of-the-art performance on 22 out of 28 benchmarks compared to DeepSeek-V3
Efficient training and inference infrastructure with high MFU (47%)
Supports multiple hardware platforms through PaddlePaddle framework
Comprehensive multimodal capabilities including text and vision understanding
Includes development toolkits for resource-efficient workflows

Cons

Large model sizes (424B total parameters for largest MoE) require significant computational resources
Only one small dense variant (0.3B) available; most models are MoE with high memory footprint
Reliance on Baidu's PaddlePaddle framework may limit ecosystem compatibility for some users

Best For

General-purpose language understanding and generationVisual-language understanding and reasoning (thinking and non-thinking modes)Cross-modal reasoning and multimodal information captureInstruction following and world knowledge memorizationResource-efficient multimodal research and development