ERNIE 5.0

What it is

ERNIE 5.0 is a foundation model from Baidu, billed as the latest generation of the ERNIE series after ERNIE 4.5 and ERNIE 4.5 Turbo. A foundation model is a large model trained on broad data that can be adapted to many tasks. ERNIE 5.0 is described as natively omni-modal, meaning it was built to handle text, images, audio and video jointly rather than adding those abilities one at a time. Baidu describes it as a unified multimodal foundation model with 2.4 trillion parameters.^[1][2]

Origins

Baidu unveiled ERNIE 5.0 at Baidu World 2025 on 13 November 2025. At the same time it opened a preview to the public on ERNIE Bot and to enterprises on Qianfan.^[1]

The technical report, posted to arXiv (arXiv 2602.04705) on 4 February 2026, describes an ultra-sparse mixture-of-experts design with modality-agnostic expert routing. The model was trained from scratch on all modalities with a unified next-group-of-tokens prediction objective.^[3] The authors describe it as the first production-scale trillion-parameter unified autoregressive model among publicly disclosed models that support both multimodal understanding and generation.^[3]

The preview ended with an updated version, ERNIE-5.0-0110, which exited preview in January 2026. On the LMArena text leaderboard, a public ranking in which people compare model answers, it reached 1,460 points and 8th place, the only Chinese model in the top 10.^[6][7] Baidu announced enterprise and consumer availability on 22 January 2026. The model is proprietary, meaning its weights are not released publicly.^[8]

Architecture and training

According to the technical report, the model's activation rate is below 3%, so only a small fraction of its parameters is used for any given token. Elastic training produces sub-models from the same network. One such sub-model uses 35.8% of the total parameters and 53.7% of the activated parameters while keeping near-full performance.^[4]

Context was extended from 8K to 32K and then to 128K during mid-training. The model was trained on PaddlePaddle, Baidu's deep-learning framework.^[4]

Reported results

Baidu's report gives post-training scores for a range of tests. On AIME 2025, a set of competition mathematics problems, the model scored 89.06. It scored 83.80 on MMLU-Pro, a harder version of a broad knowledge test, and 86.36 on GPQA-Diamond, a set of graduate-level science questions. It scored 76.21 on LiveCodeBench v6, a coding test built from recent contest problems, and 25.81 on HLE (Humanity's Last Exam). On the multimodal tests it scored 68.63 on MMMU-Pro, 84.80 on MathVista (visual mathematical reasoning) and 95.45 on DocVQA (document question answering), and 90.1 on GenEval (image generation).^[4]

Baidu cited strengths in multimodal understanding, instruction following, creative writing, factual reasoning, agentic planning and tool use.^[1]

DeepLearning.AI reported that the model outputs text, image, audio and video of up to 64,000 tokens, and that Qianfan prices it at $0.85 per million input tokens and $3.40 per million output tokens.^[5]

Place in the series

ERNIE 5.0 is the base for later work in the series. In May 2026 ERNIE 5.1 was built on ERNIE 5.0's pre-training foundation.^[1][9] The version history therefore runs from ERNIE 4.5 to ERNIE 5.0 and then to ERNIE 5.1, each step building on the one before.