本頁閱讀時間 10 分鐘
概述
這之所以重要,是因為它是中國運行時間最長的基礎模型系列之一,且自 2025 年年中以來,已有十個 ERNIE 4.5 模型以 Apache 2.0 授權發布。
重點摘要
- ERNIE 始於 2019 年 4 月,作為一種 BERT 風格的模型,透過遮蔽整個實體和短語,將知識注入到中文語言理解中。
- ERNIE 3.0 擁有 100 億個參數,在 4 TB 的文本加上知識圖譜上進行訓練,並於 2021 年 7 月在 SuperGLUE 榜單上名列前茅。
- 百度於 2023 年 3 月 16 日推出了 ERNIE Bot 聊天機器人,並於 2025 年 3 月 16 日向個人用戶免費開放。
- 百度於 2025 年年中在 Apache 2.0 下開源了 10 個 ERNIE 4.5 模型,其中包括一個總參數為 3000 億、活躍參數為 470 億的旗艦級混合專家模型。
- 於 2025 年 11 月 13 日在百度世界大會上預覽的 ERNIE 5.0 是一款專有全模態模型,而非開放權重版本。
深入探討
ERNIE is the name Baidu uses for its family of large AI models. The name started as an acronym, Enhanced Representation through kNowledge IntEgration, and today covers research language models, large multimodal models, and the consumer chatbot ERNIE Bot. When someone says "ERNIE" without a version number, it is worth asking which one they mean, because a 2019 ERNIE model and a 2025 ERNIE model share a brand but little else in size, design or capability.
The first ERNIE paper, posted to arXiv on April 19, 2019 by Yu Sun and colleagues at Baidu, changed one part of BERT's recipe. BERT, Google's 2018 encoder model, learns by hiding random word pieces in a sentence and predicting them. ERNIE instead hid whole units of meaning. Entity-level masking hides a full named entity such as a person or place, which usually spans several tokens, and phrase-level masking hides a multi-word phrase that acts as one concept. To fill those gaps, the model has to use what it has learned about the world rather than guessing from neighboring characters. The authors reported state-of-the-art results on five Chinese tasks: natural language inference, semantic similarity, named entity recognition, sentiment analysis and question answering, plus better scores on a cloze test of factual knowledge.
ERNIE 2.0, posted in July 2019 and accepted at AAAI 2020, kept the idea of richer training signals but changed how they were added. It used continual multi-task pretraining: new pretraining tasks were introduced over time and learned alongside earlier ones so the model did not forget them. The tasks targeted lexical, syntactic and semantic information, such as named entities, how close two sentences are in meaning, and discourse relations between sentences. The paper reports that ERNIE 2.0 beat BERT and XLNet across 16 tasks, including the English GLUE benchmark and several Chinese datasets.
ERNIE 3.0, posted on July 5, 2021, scaled the approach to 10 billion parameters trained on a 4 TB corpus that combined plain text with a large knowledge graph. It also merged two families of design: an auto-encoding network, which reads text in both directions and is good at understanding, and an auto-regressive network, which predicts text left to right and is good at generation. The paper reports better results than prior models on 54 Chinese tasks, and says its English version ranked first on the SuperGLUE leaderboard on July 3, 2021 with a score of 90.6, above the 89.8 human baseline listed on that leaderboard.
The consumer era began on March 16, 2023, when Baidu presented ERNIE Bot in Beijing. The Associated Press reported that the demonstration used pre-recorded video rather than a live session, that CEO Robin Li said 650 companies had signed up, and that Baidu's Hong Kong shares fell as much as 10% during the presentation before closing down 6.4%. Li said the product was not yet perfected but was being released because of demand. Two years later, on March 16, 2025, Baidu announced ERNIE 4.5, described as a native multimodal foundation model for text, images, audio and video, and ERNIE X1, a reasoning model that can call tools such as search, document question answering and a code interpreter. The same release made ERNIE Bot free to individual users ahead of a planned April 1 date.
The largest change in strategy came at the start of July 2025, when TechNode reported that Baidu had open-sourced 10 ERNIE 4.5 models on Hugging Face, GitHub and its PaddlePaddle platform under the Apache 2.0 license, which allows research and commercial use subject to its terms. The flagship text model, ERNIE-4.5-300B-A47B, is a mixture-of-experts (MoE) model. An MoE model splits parts of the network into many smaller "experts" and uses a small router to send each token to only a few of them, so a model can hold many parameters while spending compute on only some of them. According to the model card, this model has 300 billion total parameters, 47 billion active per token, 54 layers, 64 text experts with 8 active per token, and a context window of 131,072 tokens. The broader family, per TechNode, ranges up to 424 billion total parameters and includes smaller MoE models with 3 billion active parameters and 0.3 billion parameter dense models; Baidu also reported training efficiency of up to 47% model FLOPs utilization. On November 13, 2025, at Baidu World 2025, Baidu previewed ERNIE 5.0, a proprietary model it describes as natively omni-modal, available in preview through ERNIE Bot and the Qianfan platform, according to eWeek.
Compared with alternatives, ERNIE's distinctive traits are its long focus on Chinese language tasks, its tight integration with Baidu's own stack (the PaddlePaddle deep learning framework, Baidu Search, and Baidu AI Cloud), and a mixed release policy in which some generations are open-weight and others are closed. Readers comparing it with Western models such as GPT, Claude or Gemini, or with other Chinese families such as Qwen and DeepSeek, should rely on independent benchmarks rather than launch claims. Baidu's March 2025 release said ERNIE 4.5 outperformed GPT-4.5 on multiple benchmarks without naming them in the release, and such statements are company claims rather than verified results.
There are practical limits to keep in mind. ERNIE Bot is built for the Chinese market and operates under Chinese content rules, so answers on politically sensitive topics can differ from those of models built elsewhere. Version names are easy to confuse: ERNIE, ERNIE Bot, ERNIE 4.5 and ERNIE X1 refer to different things. Open weights cover specific ERNIE 4.5 checkpoints, not every model Baidu serves, and ERNIE 5.0 was announced as proprietary. Like all large language models, ERNIE models can produce confident but wrong statements, so outputs that matter should be checked.
In practice there are three ways to use ERNIE. Individuals can use ERNIE Bot for free. Businesses can pay per token through Qianfan, where Baidu priced ERNIE X1 from RMB 0.002 per thousand input tokens and RMB 0.008 per thousand output tokens at launch. Teams that want control can self-host the open ERNIE 4.5 checkpoints, but the flagship needs serious hardware: the model card lists 16 GPUs with 80 GB each in bf16, or 8 such GPUs with FP8 quantization, while the smaller 3 billion active parameter models fit far more modest setups.
04工作範例
在使用前追蹤單個 ERNIE 檢查點的功能
開啟 Hugging Face 頁面以查看 baidu/ERNIE-4.5-300B-A47B-PT,並留意授權欄位,顯示為 Apache 2.0。
閱讀配置表並記錄總參數 (300B)、主動參數 (47B)、層數 (54) 和上下文長度 (131,072 個 Token)。
估算記憶體:3000 億個權重以 bf16 格式儲存,每個權重佔用 2 位元組,總計約 600 GB,這解釋了模型卡建議使用 16 張 80 GB GPU 的原因。
與 FP8 選項進行比較:每個權重約佔用 1 位元組,記憶體需求減半至約 300 GB,符合 8 張 80 GB GPU 的配置。
檢查任務是否需要圖像;如果需要,請改為查看 ERNIE-4.5-VL 檢查點,因為 300B-A47B 模型僅支援文字。
如果硬體不足,請比較千帆 API 每千個 Token 的成本與租用 GPU 進行自行託管的費用。
它顯示了什麼
撰寫一份簡短的決策說明,陳述哪個 ERNIE 檢查點適合該任務、其授權是否允許預期用途,以及需要多少硬體或 API 支出。請根據模型卡的數據進行核對;一旦計入啟用參數和快取,記憶體估算值應在模型卡 GPU 建議值的 20% 誤差範圍內。
現實世界的實施
研究預訓練目標的研究人員將 2019 年 4 月 arXiv 論文中所述的 ERNIE 1.0 實體層級與片語層級遮罩,與 BERT 在中文任務(例如命名實體辨識與情感分析)上的隨機單字遮罩進行了比較。
中國的個人用戶可以透過百度於 2025 年 3 月 16 日對個人免費開放的 ERNIE Bot 網站,免費與 ERNIE 4.5 以及推理模型 ERNIE X1 進行對話。
企業開發人員透過百度智能雲的千帆模型即服務平台調用 ERNIE 4.5,百度在 2025 年 3 月的定價為每千個輸入 Token 0.004 人民幣、每千個輸出 Token 0.016 人民幣。
根據模型卡片,需要自架模型的團隊在 Apache 2.0 授權下從 Hugging Face 下載 ERNIE-4.5-300B-A47B,並使用 16 張 80 GB 的 GPU 以 bf16 格式,或使用 8 張配備 FP8 量化的 GPU,透過 vLLM 來提供服務。
根據 TechNode 在 2025 年 7 月開源發布時所報導的兩項配套工具,將開源的 ERNIE 4.5 模型適配於自身資料的團隊會使用百度的 ERNIEKit 工具包進行微調,並使用 FastDeploy 跨硬體進行部署。
百度 ERNIE 的未來:模型系列、ERNIE Bot 與開放權重
Several open questions shape where the ERNIE line goes next. The first is the balance between open and closed releases. Baidu published ERNIE 4.5 weights under Apache 2.0 in mid-2025, then announced ERNIE 5.0 in November 2025 as a proprietary, natively omni-modal model offered through ERNIE Bot and the Qianfan cloud platform, so it is not settled which future generations, if any, will ship open weights. The second is hardware. At the same November 2025 event, Baidu announced its Kunlunxin M100 and M300 AI chips, scheduled by the company for 2026 and 2027 respectively, according to eWeek. Whether future ERNIE models train and serve mainly on Baidu's own chips is a practical question for cost and supply. The third is evaluation. Baidu's launch materials compare ERNIE with GPT-4.5 and DeepSeek R1, but independent, reproducible tests in both Chinese and English remain the better guide, and the open ERNIE 4.5 checkpoints make such testing possible for anyone with the hardware. Research directions visible in Baidu's own descriptions include reasoning models that use tools, joint training across text, image, audio and video, and low-bit quantization for cheaper serving.
付諸實踐
為什麼這很重要
- 供應商策略. 供應商路線圖會影響您的團隊接下來可以建立的功能。
- 成本與預算. 商業條款和部署選項會影響長期成本和風險。
- 風險與安全. 公司激勵措施塑造了產品預設、安全態勢和開放性。
留意
- 發佈公告可能會超過實際生產工作流程的穩定性。
- API 定價或政策轉變可能會在一夜之間打破假設。
- 單一供應商依賴性增加了鎖定和遷移成本。
後續步驟
- 使用您自己的任務和資料集評估提供者。
- 在整合之前查看隱私、安全和法律條款。
- 維護跨模型或供應商的後備計劃。
- 監控發行說明,以便路線圖的變更不會讓團隊感到意外。
資料來源與延伸閱讀
- arXivERNIE: Enhanced Representation through Knowledge Integration
- arXiv (AAAI 2020)ERNIE 2.0: A Continual Pre-training Framework for Language Understanding
- arXivERNIE 3.0: Large-scale Knowledge Enhanced Pre-training for Language Understanding and Generation
- Associated Press (via News4JAX)百度發表 ChatGPT 競爭對手文心一言;650 家企業註冊使用
- Baidu press release via PR Newswire (Taiwan News)百度發表 ERNIE 4.5 及推理模型 ERNIE X1,並提前宣佈 ERNIE Bot 免費
- TechNode百度開源 ERNIE 4.5 系列模型,包含多模態 MoE 架構
- Hugging Face (Baidu)baidu/ERNIE-4.5-300B-A47B-PT 模型卡片
- eWeek百度發表 ERNIE 5.0 與新型 AI 晶片
不斷探索
常見問題
什麼是百度 ERNIE:模型家族、ERNIE Bot 與開源權重?
ERNIE 是百度的 AI 模型家族,於 2019 年作為 BERT 的知識增強替代方案起步,並發展成為支援 ERNIE Bot 聊天機器人的大型多模態模型。它之所以重要,是因為它是中國歷史最悠久的基礎模型系列之一,且自 2025 年中以來,已有十個 ERNIE 4.5 模型以 Apache 2.0 授權條款發布。
哪家公司開發了 ERNIE 系列型號?
2019 年的 ERNIE 論文列出了百度作者的所屬機構。這標識了該論文背后的公司,但並未宣稱百度是排名第一或「領先」的搜尋引擎。
在最初的 2019 年 ERNIE 模型中,知識遮罩(knowledge masking)與 BERT 的 Token 遮罩有何不同?
原始論文在詞級學習中加入了實體級與片語級的遮罩,允許對跨多個 Token 的語義範圍進行訓練。
百度將其於 2023 年公布的對話助理命名為何?
百度的 2023 年公告將該對話產品命名為 ERNIE Bot(文心一言)。
縮寫 ERNIE 主要強調什麼?
ERNIE 代表Enhanced Representation through kNowledge IntEgration,突顯了其知識重點。
最初的 ERNIE 論文回報了多少個中文自然語言處理任務的結果?
最初的 ERNIE 論文回報了五個中文 NLP 任務的結果:自然語言推論、語義相似度、命名實體識別、情感分析與問答。