Baidu develops the ERNIE model family, ranging from dense to multimodal MoE architectures (up to 2.4T parameters in ERNIE 5.0), trained on PaddlePaddle using FP8 mixed-precision and heterogeneous hybrid parallelism. Tooling includes ERNIEKit for SFT, LoRA, and DPO, and FastDeploy for inference with vLLM and OpenAI API compatibility. Also operates Apollo Go autonomous driving.
Baidu launched a neutral marketplace for AI model services with 100+ models processing 700B tokens daily, aiming to standardize how enterprises buy inference.
Baidu's new automated research system modifies code, runs experiments, and keeps only verified wins, reaching SOTA on NanoGPT Speedrun and NanoChat AutoResearch.
Baidu's Unlimited-OCR uses a novel attention mechanism to parse entire multi-page documents in a single pass, eliminating the chunking bottleneck that plagues current OCR pipelines.
Baidu's new 4B parameter open-weight vision-language model tops OmniDocBench v1.5 and OCRBench while unifying parsing, layout, and understanding in one pass.
Baidu Search unveils a four-agent blueprint that plans, executes, and reflects on multi-step queries, moving beyond one-shot retrieve-then-generate RAG.
Baidu's compact 21B Mixture-of-Experts model activates only 3B parameters per token, targeting deep reasoning, tool use, and 128K context on a single GPU.