ERNIE 5.0 Technical Report

In this report, we introduce ERNIE 5.0, a natively autoregressive foundation model desinged for unified multimodal understanding and generation across text, image, video, and audio. All modalities are trained from scratch under a unified next-group-of-tokens prediction objective, based on an ultra-sparse mixture-of-experts (MoE) architecture with modality-agnostic expert routing. To address practical challenges in large-scale deployment under diverse resource constraints, ERNIE 5.0 adopts a novel elastic training paradigm. Within a single pre-training run, the model learns a family of sub-models with varying depths, expert capacities, and routing sparsity, enabling flexible trade-offs among performance, model size, and inference latency in memory- or time-constrained scenarios. Moreover, we systematically address the challenges of scaling reinforcement learning to unified foundation models, thereby guaranteeing efficient and stable post-training under ultra-sparse MoE architectures and diverse multimodal settings. Extensive experiments demonstrate that ERNIE 5.0 achieves strong and balanced performance across multiple modalities. To the best of our knowledge, among publicly disclosed models, ERNIE 5.0 represents the first production-scale realization of a trillion-parameter unified autoregressive model that supports both multimodal understanding and generation. To facilitate further research, we present detailed visualizations of modality-agnostic expert routing in the unified model, alongside comprehensive empirical analysis of elastic training, aiming to offer profound insights to the community.

XGC-VQA: A unified video quality assessment model for User, Professionally, and Occupationally-Generated Content

With the rapid growth of Internet video data amounts and types, a unified Video Quality Assessment (VQA) is needed to inspire video communication with perceptual quality. To meet the real-time and universal requirements in providing such inspiration, this study proposes a VQA model from a classification of User Generated Content (UGC), Professionally Generated Content (PGC), and Occupationally Generated Content (OGC). In the time domain, this study utilizes non-uniform sampling, as each content type has varying temporal importance based on its perceptual quality. In the spatial domain, centralized downsampling is performed before the VQA process by utilizing a patch splicing/sampling mechanism to lower complexity for real-time assessment. The experimental results demonstrate that the proposed method achieves a median correlation of $0.7$ while limiting the computation time below 5s for three content types, which ensures that the communication experience of UGC, PGC, and OGC can be optimized altogether.