Baidu

A training-infrastructure method for the load imbalance MoE's dynamic routing causes in expert-parallel training. Prior dynamic-replica schemes copy h…

Baidu's one-shot long-horizon document parsing model, positioned to "push DeepSeek-OCR one step further": a compact DeepSeek-OCR-derived MoE decoder (…

Successor to ERNIE 5.0, launched at Baidu Create 2026. Derived from ERNIE 5.0 via elastic sub-network extraction — total params ~1/3 of 5.0, act…

Technical paper on a 4B parameter vision-language model using Layout-as-Thought mechanisms.

Introduced Reinforcement Learning from Community Feedback (RLCF) for aligning AI with scientific reasoning, accompanied by a dataset of 700,000 scient…

Most advanced ERNIE model to date. 2.4 trillion total parameters with ultra-sparse MoE (<3% activation ratio) for extreme inference efficiency. Native…

An analysis of why some model families benefit dramatically from reinforcement learning while others barely move. The authors identify "distributional…

Kernel library for high-speed LLM serving on PaddlePaddle.

Newest inference-focused chip designed for low-power, high-throughput LLM deployment.

Family of 10 open-source multimodal MoE models ranging from 424B total parameters (47B active) down to a 0.3B dense model. Uses a heterogeneous modali…

Deep-thinking reasoning model series with multimodal capabilities. ERNIE X1 delivers performance on par with DeepSeek-R1 at lower cost. ERNIE X1.1 (Se…

Proposes AdaGC, an adaptive per-tensor gradient clipping method that bounds each tensor's gradient norm relative to an exponential moving average of i…

Optimized version for enterprise use cases requiring higher speed and lower costs.

Benchmark for open-domain dialogue quality assessment.

Significantly improved reasoning, memory, and generation capabilities.

Integrated with ERNIE Bot (Wenxin Yiyan). At release, claimed performance superior to ChatGPT (GPT-3.5) and competitive with GPT-4.

Baidu AI Cloud's one-stop model-as-a-service platform for developers to fine-tune and deploy ERNIE models.

10B parameter text-to-image model for high-quality image generation from Chinese prompts.

Baidu's first massive-scale model (260B parameters) combining structured knowledge graphs with unstructured text.

High-quality Chinese translation and localization of the popular MS-MARCO dataset.

Library for SOTA NLP models, including ERNIEKit and PaddleFormers.

Specialized for ASR, TTS, and speaker verification.

Industry standard open-source OCR toolkit, currently at version 4.0.

Baidu's proprietary AI chips (R200/R300/P800) optimized for PaddlePaddle.

End-to-end toolkit for object detection (YOLO, Faster R-CNN).

Image segmentation toolkit.

Continual multi-task learning framework that set new SOTA records on the GLUE and SuperGLUE benchmarks.

Foundational ERNIE model introducing knowledge integration into language model pretraining, surpassing BERT on Chinese NLP tasks.

Massive dataset for autonomous driving including 3D semantic segmentation and trajectory prediction.

Largest Chinese machine reading comprehension (MRC) dataset, containing real user queries.

World's largest open-source autonomous driving platform.

China's first and most widely used domestic deep learning framework. Core engine supporting multi-device (CPU, GPU, XPU) training and inference.

Landmark papers in end-to-end deep learning for speech recognition.