ERNIE (Baidu)

ERNIE (Baidu model family) is a family of pretrained language models and multimodal foundation models developed by Baidu since 2019. The name originally expanded to Enhanced Representation through kNowledge IntEgration in the 2019 ERNIE language-representation paper; it was subsequently retained as the brand for Baidu's broader foundation-model line.

The technical meaning of “knowledge integration” has changed across generations. ERNIE 1.0 learned semantic units indirectly through phrase-level and entity-level masking rather than by inserting a complete external knowledge graph into every input. ERNIE 2.0 added continual multi-task pre-training, while ERNIE 3.0 and ERNIE 3.0 Titan introduced explicit joint objectives over text and knowledge-graph triples. Later generations shifted toward large-scale generative models, Mixture-of-Experts (MoE), multimodality, reasoning, tool use, and agentic post-training.

As of 10 July 2026, Baidu's Qianfan model catalogue describes ERNIE 5.1 as the latest general-purpose ERNIE flagship, with particular emphasis on agents, knowledge, reasoning, and deep search. The same catalogue continues to expose ERNIE 5.0 as the native omni-modal model for text, image, video, and audio tasks, and ERNIE X1.1 as a dedicated reasoning branch. The hosted ERNIE 5.1 and ERNIE 5.0 text endpoints have a 128,000-token context window, a maximum listed input of 119,000 tokens, and a maximum listed output of 65,536 tokens.

Baidu released the ten-model ERNIE 4.5 family with downloadable weights and development tooling under the Apache License 2.0 in June 2025. It later released other open-weight specialist models, including the 8-billion-parameter text-to-image model ERNIE-Image. By contrast, Baidu's official model repositories did not list an ERNIE 5.0 or ERNIE 5.1 weight release as of 10 July 2026; those models were documented primarily through ERNIE Bot and the hosted Qianfan platform.

History and background

Development of the ERNIE line began in Baidu's research organization in 2019, following the success of BERT and related contextual language models. Chinese text presents particular tokenization and semantic-unit challenges because words are not ordinarily separated by spaces and many meaningful entities and phrases contain several characters. The early ERNIE work therefore focused on masking complete semantic units rather than isolated tokens.

ERNIE 1.0 (2019)

The first ERNIE paper was submitted in April 2019. It introduced three masking strategies for a BERT-like encoder: basic token masking, phrase-level masking, and entity-level masking. Phrase and entity boundaries were supplied by lexical analysis and named-entity recognition. This encouraged the model to recover an entire semantic unit from its wider context.

ERNIE 1.0 used a 12-layer Transformer encoder with a hidden size of 768 and 12 attention heads, broadly corresponding to the BERT-base scale. Its Chinese pre-training data comprised approximately 173 million sentences: about 21 million from Chinese Wikipedia, 51 million from Baidu Baike, 47 million from news, and 54 million dialogue turns from Baidu Tieba. A Dialogue Language Model objective was used for multi-turn conversational data.

The authors reported state-of-the-art results on five Chinese-language tasks at publication time. Selected reported test results included 78.4 accuracy on the Chinese XNLI set, compared with 77.2 for the paper's BERT baseline, and 93.8 F1 on MSRA-NER, compared with 92.6 for BERT. These figures reflect the authors' 2019 evaluation setup and should not be treated as current model comparisons.

ERNIE 2.0 (2019)

ERNIE 2.0, released as a preprint in July 2019 and presented at AAAI 2020, introduced a continual multi-task pre-training framework. Instead of replacing the complete training objective whenever a new task was introduced, the framework incrementally added objectives while continuing to sample earlier ones, reducing catastrophic forgetting.

The seven principal objectives were grouped as follows:

• Word-aware tasks - knowledge masking, capitalization prediction, and token-document relation prediction.
• Structure-aware tasks - sentence reordering and sentence-distance prediction.
• Semantic-aware tasks - discourse-relation prediction and information-retrieval relevance prediction.

ERNIE 2.0 was trained in Chinese and English using encyclopedias, books, news, dialogue or forum text, and search-related data. The paper reported a GLUE development-set average of 80.6 for the base model and a GLUE test-set score of 83.6 for the large model. Because the two numbers refer to different model sizes and evaluation splits, they should not be read as a controlled base-versus-large comparison. The authors reported improvements over their BERT and XLNet baselines across 16 English and Chinese tasks.

ERNIE 3.0 (2021)

ERNIE 3.0, published in July 2021, combined auto-encoding and autoregressive pre-training in a unified framework. A Universal Representation Module learned shared lexical, syntactic, and semantic features, while task-specific representation modules used different attention masks and recurrent-memory settings for natural-language understanding and generation.

The 10-billion-parameter model was trained on approximately 4 TB of Chinese text in 11 categories, including encyclopedias, web pages, search logs, question-answer data, and legal, financial, medical, literary, and conversational material. The knowledge component contained more than 50 million facts. The paper reported state-of-the-art results on 54 Chinese NLP tasks and a SuperGLUE score of 90.6 as of 3 July 2021, compared with the benchmark's then-published human baseline of 89.8. The result was a developer-reported snapshot under the evaluation rules and model comparisons available at that date.

ERNIE 3.0 Titan (2021)

ERNIE 3.0 Titan scaled the framework to a 260-billion-parameter dense model implemented in PaddlePaddle. Its December 2021 report added objectives for credible and controllable generation and described online distillation for producing smaller student models during pre-training.

Titan's Universal Representation Module contained 48 layers, a hidden size of 12,288, 192 attention heads, and a feed-forward inner dimension of 196,608. Separate 12-layer task-specific modules with a hidden size of 768 were used for natural-language understanding and generation. The authors evaluated the system across 68 datasets and reported improvements over the comparison systems used in the paper. Independent reproduction was limited because the full 260-billion-parameter checkpoint, data pipeline, and training infrastructure were not released.

ERNIE-ViLG and early multimodal work

In parallel with the main language-model line, Baidu developed multimodal ERNIE systems. ERNIE-ViLG 2.0, published in 2022, was a text-to-image diffusion model using a knowledge-enhanced mixture of denoising experts. It was an image-generation model rather than a direct predecessor with the same architecture as the later ERNIE 4.5 or ERNIE 5.0 foundation models.

ERNIE Bot and commercial generations (2023-2024)

Baidu demonstrated ERNIE Bot (Wenxin Yiyan, 文心一言) on 16 March 2023. The initial event did not constitute unrestricted public availability; general public access opened on 31 August 2023 after Chinese regulators approved the first group of public generative-AI services.

Baidu launched ERNIE 3.5 in May 2023 as the foundation model then powering ERNIE Bot, not in June as sometimes reported. ERNIE 4.0 followed in October 2023 and was made available in ERNIE Bot and through Baidu's cloud API. ERNIE 4.0 Turbo, released in June 2024, was positioned as faster and less expensive than ERNIE 4.0. Baidu did not publish a technical report establishing the frequently repeated claim that it generated “more than one thousand words in twenty seconds”, so that claim is omitted here.

Baidu reported 200 million ERNIE Bot users in April 2024 and 300 million by June 2024. These were cumulative user figures announced by the company, not monthly active-user counts. ERNIE Bot became free for desktop and mobile users on 1 April 2025.

ERNIE 4.5 and ERNIE X1 (2025)

Baidu launched the hosted ERNIE 4.5 and the reasoning-oriented ERNIE X1 in March 2025, followed by ERNIE 4.5 Turbo and ERNIE X1 Turbo in April. On 30 June 2025, Baidu released ten ERNIE 4.5 checkpoints with weights and supporting code under the Apache License 2.0.

The initial open-weight family comprised text-only and vision-language variants at several scales: 300B-A47B and 21B-A3B text MoE models, 424B-A47B and 28B-A3B vision-language MoE models, and a 0.3B dense text model, with base and post-trained checkpoints. In the naming convention, the number after “A” denotes approximately activated parameters per token. The published checkpoints support a 131,072-token context window.

Baidu later published thinking-oriented ERNIE 4.5 checkpoints, including ERNIE-4.5-21B-A3B-Thinking and ERNIE-4.5-VL-28B-A3B-Thinking. These were additions to, rather than members of, the original set of ten June 2025 checkpoints.

ERNIE X1.1 was introduced in September 2025 as an updated reasoning model. Baidu's current Qianfan catalogue lists the production ERNIE-X1.1 endpoint with a 64,000-token context, a maximum input of 55,000 tokens, and a combined reasoning-plus-answer output limit of 65,536 tokens.

ERNIE 5.0 (2025-2026)

Baidu unveiled ERNIE 5.0 in November 2025 and released an updated version in January 2026. The detailed technical report was submitted on 4 February 2026, and Baidu published its English technical overview on 6 February.

ERNIE 5.0 is a 2.4-trillion-parameter native omni-modal model trained from scratch to perform both understanding and generation across text, images, video, and audio in a unified autoregressive framework. It uses a shared token space, a unified Next-Group-of-Tokens Prediction objective, an ultra-sparse MoE backbone with modality-agnostic routing, and elastic training across model depth, expert-pool width, and routing sparsity. Baidu reports that fewer than 3% of all parameters are activated for each token.

ERNIE 5.1 and ERNIE-Image (2026)

Baidu officially released ERNIE 5.1 on 9 May 2026. It inherits the pre-training foundation of ERNIE 5.0 but extracts a more compact configuration from the elastic super-network. Baidu states that ERNIE 5.1 has approximately one-third as many total parameters and approximately one-half as many activated parameters as ERNIE 5.0. The company has not disclosed an exact absolute parameter count; figures such as “800 billion parameters” are therefore estimates rather than an official specification.

ERNIE 5.1 also introduced a disaggregated, fully asynchronous reinforcement-learning infrastructure that separates training, inference, reward computation, and agent-loop control. Baidu describes scaled agentic post-training as the main source of improvements in long-horizon tool use, knowledge, reasoning, and deep search. The model is the latest ERNIE text and agent flagship in the July 2026 Qianfan catalogue, while ERNIE 5.0 remains the listed native omni-modal endpoint for visual understanding.

ERNIE-Image, released with open weights on 15 April 2026, is a separate 8-billion-parameter, single-stream Diffusion Transformer for text-to-image generation. ERNIE-Image-Turbo uses the same general model line with an accelerated eight-step inference configuration. These image models should not be confused with the 2.4-trillion-parameter ERNIE 5.0 omni-modal foundation model.

Theoretical foundations and architectural principles

Knowledge masking

The defining mechanism of the first ERNIE model was not direct graph lookup but knowledge-oriented masking. Standard masked language modelling usually hides individual tokens. ERNIE 1.0 additionally hides every token belonging to a recognised phrase or named entity. For a semantic unit $E=\{t_1,\dots ,t_k\}$ and a selected mask set $ℳ$, a simplified objective is:

${ℒ}_{\mathrm{M}\mathrm{L}\mathrm{M}}=-\sum _{i\in ℳ}\log P_{\theta }(t_i\mid x_{\setminus ℳ}).$

Masking the complete unit forces the encoder to reconstruct it from the surrounding sentence rather than from the visible characters inside the unit. The phrase “knowledge integration” in ERNIE 1.0 therefore refers chiefly to injecting linguistic boundaries into the corruption process. It should not be described as supplying a structured knowledge graph to every training example.

ERNIE 2.0 generalized this approach through continual multi-task pre-training. New lexical, structural, and semantic objectives were introduced incrementally while earlier objectives continued to be sampled. The method was intended to retain previously learned capabilities while expanding the set of pre-training signals.

Unified ERNIE 3.0 framework

ERNIE 3.0 separated the network conceptually into two levels:

• Universal Representation Module — a shared Transformer-XL-based module trained across the full collection of pre-training tasks.
• Task-specific Representation Modules — additional modules and attention patterns for bidirectional understanding and left-to-right generation.

In ERNIE 3.0 Titan, the Universal Representation Module used 48 layers, a hidden size of 12,288, 192 attention heads, and a feed-forward dimension of 196,608. The smaller task-specific modules used 12 layers, a hidden size of 768, and 12 attention heads. The combined design allowed auto-encoding objectives to support natural-language understanding and autoregressive objectives to support generation without training two completely separate foundation models.

Universal Knowledge-Text Prediction

Universal Knowledge-Text Prediction (UKTP) connected encyclopedia text with knowledge-graph triples. A training item associated text $x$ with a triple $(h,r,t)$, where $h$ and $t$ are the head and tail entities and $r$ is their relation. One direction predicted the masked relation from the text and entity representations:

$P_{\theta }(r\mid x,h,t)=\mathrm{softmax}(W{f}_{\theta }(x,h,t)+b).$

The complementary direction masked text tokens and used the triple as additional evidence. UKTP is the clearest point in the early ERNIE line at which explicit structured facts, rather than only phrase and entity boundaries, became a direct part of the pre-training objective.

Credible and controllable generation in Titan

ERNIE 3.0 Titan added two generation-oriented objectives.

Self-supervised adversarial loss. Original paragraphs were treated as positive examples, while continuations generated by an earlier ERNIE model from one-to-three-sentence prefixes were treated as negative examples. A simplified discriminator objective is:

${ℒ}_{\mathrm{a}\mathrm{d}\mathrm{v}}=-\sum _n[y_n\log D_{\theta }(x_n)+(1-y_n)\log (1-D_{\theta }(x_n))],$

where $y_n=1$ denotes an original paragraph. The objective was intended to improve the model's ability to distinguish human corpus text from model-generated text. It is not a proof that Titan generations are factually reliable, because the negative distribution was limited by the earlier generator used to create it.

Controllable language-modelling loss. Examples were annotated with attributes such as genre, topic, keywords, sentiment, and target length. Conditional examples optimized:

${ℒ}_{\mathrm{c}\mathrm{t}\mathrm{r}\mathrm{l}}=-\sum _t\log P_{\theta }(x_t\mid x_{<t},c),$

where $c$ is the attribute prompt. Titan alternated conditional and unconditional language-modelling examples so that generation remained possible without an attribute prompt.

Heterogeneous multimodal MoE in ERNIE 4.5

ERNIE 4.5 moved from the dense Titan architecture to heterogeneous multimodal MoE. Its vision-language variants contain separate text and vision experts, and routing is isolated by modality. Baidu introduced a router-orthogonalization regularizer to encourage different routing projections to specialize rather than collapse toward the same expert choices, together with token-balancing mechanisms for stable expert utilization.

A simplified MoE transformation for token representation $𝐡$ is:

$\mathrm{MoE}(𝐡)=\sum _{i\in 𝒮(𝐡)}{g}_i(𝐡)E_i(𝐡),$

where $𝒮(𝐡)$ is the routed expert subset and $g_i$ are normalized gate weights. In the 300B-A47B text checkpoint, the published configuration has 64 experts and activates eight per token; the model has 300 billion total parameters and approximately 47 billion activated parameters.

The multimodal training pipeline introduced a variable-resolution vision encoder, feature projection, vision experts, three-dimensional rotary position encodings for spatial and temporal positions, and adaptive frame sampling for video. The published training process was staged: text foundations, vision-specific learning, and joint multimodal training. Text-only checkpoints were subsequently extracted from the multimodal training system by retaining the text-related parameters.

FlashMask represents heterogeneous causal, bidirectional, and block-sparse attention masks more compactly during training. The report's linear-complexity claim concerns the representation and manipulation of the mask; it does not mean that every ERNIE 4.5 attention operation has linear computational complexity in sequence length.

Native omni-modality in ERNIE 5.0

ERNIE 5.0 replaces ERNIE 4.5's modality-isolated expert pools with modality-agnostic routing over a shared pool. Text, image, video, and audio inputs are tokenized, serialized into a common sequence, and processed by one autoregressive backbone. Routing depends on the unified token representation rather than on an explicit modality identifier, allowing expert specialization to emerge from task requirements.

Its Next-Group-of-Tokens Prediction framework unifies several predictive structures:

• text uses standard next-token prediction augmented with multi-token prediction;
• image and video generation use Next-Frame-and-Scale Prediction;
• audio generation uses Next-Codec Prediction.

This is a unified training objective, not a claim that raw pixels and waveforms are processed without tokenizers. ERNIE 5.0 includes separate visual and audio tokenization pipelines before the shared backbone.

The ultra-sparse MoE activates fewer than 3% of the model's parameters for a token. Sparse activation reduces per-token computation relative to a dense 2.4-trillion-parameter network, but the complete expert set must still be stored or distributed for general inference.

Elastic training and ERNIE 5.1

ERNIE 5.0 applies Once-For-All-style elastic training along three dimensions:

• Elastic depth — training alternates between the full Transformer stack and shallower subnetworks.
• Elastic width — training alternates between the full expert pool and subsets of experts.
• Elastic sparsity — the number of experts activated per token varies during training.

The full model and sampled subnetworks are optimized in the same pre-training run. Baidu reported that an experimental jointly compressed configuration retained competitive performance with 35.8% of total parameters and 53.7% of activated parameters, while reducing routing top-k to 25% produced more than a 15% decoding-speed increase with a minor average accuracy loss in its tests. These are experimental results from the technical report, not universal deployment guarantees.

ERNIE 5.1 was selected from this elastic model family and then subjected to separate post-training. Baidu reports approximately one-third of ERNIE 5.0's total parameters and one-half of its activated parameters, but has not published the exact depth, expert count, routing top-k, or absolute parameter total of the production checkpoint.

Pre-training objectives and data

ERNIE 3.0 and Titan

The ERNIE 3.0 corpus comprised approximately 4 TB of Chinese data divided into 11 broad categories, including encyclopedia articles, web text, search logs, question-answer material, dialogue, news, and legal, financial, medical, literary, and poetic writing. The accompanying knowledge graph contained more than 50 million facts.

Its task families included:

Word-aware objectives:

• phrase-level and entity-level masked language modelling;
• document language modelling with Transformer-XL recurrence.

Structure-aware objectives:

• sentence reordering;
• sentence-distance classification.

Knowledge- and generation-aware objectives:

• Universal Knowledge-Text Prediction;
• original-versus-generated paragraph discrimination;
• attribute-conditioned controllable generation.

For Titan's adversarial dataset, the authors sampled approximately two million original paragraphs and generated corresponding negative continuations with an earlier ERNIE model. The controllable-generation data included 26 topic classes, sentiment and length attributes, keywords, and learnable genre prompts; the number of genre-prompt tokens was sampled between zero and 64.

ERNIE 4.5

ERNIE 4.5 was pretrained on trillions of text and multimodal tokens. The report describes large-scale deduplication, quality filtering, synthetic-data construction, and progressively longer sequences, but it does not release a complete enumerated training corpus or sufficient information to reproduce the original data mixture exactly.

The family used staged multimodal pre-training rather than simply attaching a visual adapter after all language training had finished. The final multimodal phase jointly optimized text and visual examples, after which text-only and vision-language checkpoints could be derived from the shared training system.

ERNIE 5.0

ERNIE 5.0 was trained from scratch on trillions of tokens across text, image, video, and audio data. Baidu describes the text mixture as multilingual and multi-domain and the multimodal mixture as combining paired and interleaved data, understanding examples, and generative data. The full source inventory, licensing composition, and language proportions are not public.

Pre-training began with an 8,000-token stage and then extended the context progressively to 32,000 and 128,000 tokens during mid-training. The RoPE base was set for long-context extension from the initial stage. A posterior-based loss-weighting method rescaled losses from different modalities to reduce optimization imbalance.

Training and distributed optimization

ERNIE 3.0 Titan

Titan used the Adam optimizer with a reported peak learning rate of $10^{-4}$, ${\beta }_1=0.9$, ${\beta }_2=0.95$, weight decay of 0.1, and gradient-norm clipping at 1.0. Its main sequence length was 512 tokens, with recurrent memory of 128 tokens for generative objectives. A progressive schedule accelerated the first 4,000 optimization steps before linear learning-rate decay.

The 260-billion-parameter model was trained with four-dimensional hybrid parallelism in PaddlePaddle:

• Data parallelism replicated model partitions across data batches.
• Tensor/model parallelism split operations and parameters within layers.
• Pipeline parallelism distributed layer groups across pipeline stages.
• Group-sharded parallelism reduced duplicated optimizer and parameter states in a ZeRO-like manner.

The report describes heterogeneous NVIDIA V100 and Huawei Ascend 910 clusters and reports approximately 91.7% weak-scaling efficiency in a selected Ascend experiment. That number is specific to the authors' workload and hardware configuration.

ERNIE 4.5

Baidu reports training the largest ERNIE 4.5 system on 2,016 NVIDIA H800 GPUs with 47% Model FLOPs Utilization. The stack used heterogeneous hybrid parallelism, hierarchical load balancing for experts, FP8 mixed-precision training, recomputation, and fault-tolerant checkpointing. The report's “zero-cost checkpoint” wording refers to overlapping checkpoint work with training; it does not mean checkpoint creation consumed literally no resources.

For inference, Baidu published FastDeploy support, prefill/decode disaggregation, low-bit quantization, and multi-expert parallel collaboration. Separate CCQ research later demonstrated two-bit quantization of ERNIE-4.5-300B-A47B and single-accelerator deployment in a particular 80-GB H20 configuration. That is a specialized compressed checkpoint and engine, not the memory requirement of the standard BF16 model.

ERNIE 5.0 and 5.1

ERNIE 5.0 uses hybrid parallelism and fine-grained memory management for its ultra-sparse MoE. Visual and audio tokenizers can be placed on separate GPU nodes from the backbone so that different subsystems use different parallelization strategies. The report also describes a disaggregated reinforcement-learning infrastructure coordinating policy training, rollout inference, reward computation, and environment interaction.

For ERNIE 5.1, Baidu further separated the control plane of four subsystems—training, inference, reward, and agent loop—and allowed each to scale independently. The company also reports a unified FP8 operator stack, rollout-router replay to reduce training–inference routing divergence, networked data components, KV-cache pooling, and heterogeneous CPU/GPU resource scheduling. These descriptions are developer disclosures; detailed independent performance audits of the production training system have not been published.

Post-training and alignment

ERNIE post-training methods vary by generation and checkpoint. It is inaccurate to attribute one fixed RLHF recipe to the entire family.

ERNIE 4.5

The ERNIE 4.5 report describes supervised fine-tuning on instruction data, preference optimization, and reinforcement learning for reasoning and verifiable tasks. Baidu's open ERNIEKit supports supervised fine-tuning, parameter-efficient fine-tuning such as LoRA, and Direct Preference Optimization. Some later ERNIE 4.5 checkpoints explicitly expose thinking behavior, whereas the original June 2025 release also included conventional non-thinking instruction models.

ERNIE 5.0

ERNIE 5.0 uses supervised fine-tuning followed by Unified Multimodal Reinforcement Learning. The report describes several stabilizing components:

• an unbiased replay buffer to improve rollout reuse without intentionally skewing the task distribution;
• multi-granularity importance sampling and positive-sample masking for policy optimization;
• a unified verifier system for tasks with checkable outcomes;
• adaptive hint-based reinforcement learning for difficult sparse-reward tasks.

The hosted platform exposes both ordinary and thinking-oriented ERNIE 5.0 endpoints. The Qianfan documentation treats generated reasoning text as part of the total output allowance; it should not be described as a complete or guaranteed representation of the model's internal computation.

ERNIE 5.1

Baidu describes ERNIE 5.1 post-training as a four-stage process centered on Multi-Teacher On-Policy Distillation and scaled agentic reinforcement learning. Expert policies can be optimized in parallel for different capabilities and then integrated into a unified policy. Because the complete datasets, reward functions, teachers, and environment distribution have not been released, the contribution of each stage cannot be independently isolated.

Main model families

Model or branch	First release	Publicly disclosed scale	Architecture and modalities	Context or qualification
ERNIE 1.0	2019	BERT-base scale, approximately 110M	Dense Transformer encoder; text	512-token BERT-style encoder
ERNIE 2.0	2019	Base and large configurations, approximately BERT scale	Dense Transformer encoder; text	Continual multi-task pre-training
ERNIE 3.0	2021	10B	Unified auto-encoding/autoregressive Transformer-XL framework; text	4-TB corpus and more than 50M knowledge facts
ERNIE 3.0 Titan	2021	260B	Dense unified text model	512-token main sequence plus recurrent memory in the report
ERNIE 4.0 / 4.0 Turbo	2023 / 2024	Not publicly disclosed	Proprietary generative models; ERNIE Bot and API	Commercial hosted models
ERNIE 4.5 open family	2025	0.3B dense; MoE models from 21B-A3B to 424B-A47B	Text and vision-language checkpoints	131,072 tokens
ERNIE X1 / X1.1	2025	Not publicly disclosed	Proprietary reasoning branch; text	X1.1 endpoint: 64K context as of July 2026
ERNIE 5.0	2025 preview; 2026 technical release	2.4T total; fewer than 3% activated per token	Native autoregressive text, image, video, and audio understanding and generation	128K hosted context
ERNIE 5.1	2026	Exact total not disclosed; approximately one-third of ERNIE 5.0 total and one-half of its active parameters	General-purpose text and agent flagship derived from the ERNIE 5.0 elastic network	128K hosted context
ERNIE-Image	2026	8B	Single-stream latent Diffusion Transformer; text-to-image	Separate open-weight image-generation branch

The table mixes research checkpoints and commercial services because Baidu has used the ERNIE name for both. Parameter counts, modalities, and context limits should therefore be attached to a specific checkpoint or endpoint rather than generalized to “ERNIE” as a whole.

Reported benchmark results

Benchmark numbers below are primarily developer-reported. They depend on prompts, sampling settings, test-set versions, tool access, reasoning budgets, judge models, and baseline versions. Results from different technical reports are not a controlled time series.

Early language models

Model	Reported benchmark result	Important qualification
ERNIE 1.0	Chinese XNLI test accuracy 78.4; MSRA-NER F1 93.8	2019 paper and baselines
ERNIE 2.0 Base	GLUE development average 80.6	Development set
ERNIE 2.0 Large	GLUE test score 83.6	Different model size and split from the preceding row
ERNIE 3.0	SuperGLUE 90.6	Leaderboard snapshot reported for 3 July 2021
ERNIE 3.0 Titan	Improvements reported across 68 datasets	No single comparable aggregate score; limited independent reproduction

ERNIE 4.5

Selected results reported for the post-trained ERNIE-4.5-300B-A47B text checkpoint include:

Benchmark	Reported score	Setting or note
C-Eval	90.6	5-shot
CMMLU	90.2	Developer evaluation
MMLU	86.5	Report protocol
IFEval	88.0	Instruction following
GSM8K	91.8	Report protocol

Selected vision-language results for ERNIE-4.5-VL-424B-A47B were reported at approximately 67.3 on MMMU without an extended thinking mode and around 70 with thinking, 78.8 on MathVista in the reported thinking setup, and 883 on OCRBench. The exact checkpoint and reasoning mode matter: these multimodal scores should not be placed in a table headed only “ERNIE-4.5-300B-A47B”, which is a text checkpoint.

Baidu stated that its evaluated ERNIE 4.5 checkpoint exceeded the DeepSeek-V3 baseline used in the report on 22 of 28 selected benchmarks. This is a vendor comparison under one report's harness, not a permanent ranking against all later DeepSeek releases.

ERNIE 5.0

Selected post-trained ERNIE 5.0 results from the February 2026 report are:

Category	Benchmark	ERNIE 5.0
Knowledge	SimpleQA	74.01
Knowledge	ChineseSimpleQA	86.03
General	MMLU-Pro	83.80
STEM	GPQA-Diamond	86.36
STEM	AIME 2025	89.06
Coding	LiveCodeBench v6	76.21
Coding	HumanEval+	94.48
Instruction following	IFEval	93.35
Instruction following	Multi-IF	85.56
Agent	TAU2-Bench	78.79
Agent	ACEBench-en / ACEBench-zh	87.70 / 89.60
Agent	SpreadsheetBench	40.08

The same report showed a mixed competitive picture rather than universal dominance. For example, ERNIE 5.0 led the comparison table on SimpleQA, ChineseSimpleQA, MultiChallenge, Multi-IF, and ACEBench, while Gemini 3-Pro led several difficult mathematics, coding, and spreadsheet tasks in the authors' evaluation.

ERNIE 5.0 was also evaluated on visual understanding, image and video generation, speech recognition, audio understanding, and speech generation. Scores across these categories use different metrics and cannot be combined into a single meaningful “omnimodal accuracy” number.

ERNIE 5.1

Baidu's May 2026 release article reports strong performance on agent, knowledge, reasoning, deep-search, and creative-writing evaluations. It reports 99.6 on AIME 2026 with tools and an Arena Search score of 1,223 on 9 May 2026, corresponding at that moment to fourth globally and first among Chinese models on that leaderboard. These are date-specific developer claims. Arena positions change as models and votes are added, and several compared systems and some evaluation details are not independently documented in the release post.

Applications

ERNIE Bot, search, and document work

ERNIE Bot is Baidu's consumer conversational product. It supports writing, question answering, search-assisted research, document analysis, and multimodal interactions, with capabilities depending on the selected production model. Baidu's 200-million and 300-million figures announced in April and June 2024 were cumulative user counts rather than monthly active users. In its fourth-quarter 2025 results, Baidu separately reported that ERNIE Assistant reached 202 million monthly active users in December 2025; the two types of metric are not directly comparable.

Baidu Search uses ERNIE models for generated summaries, conversational search, content organization, and other AI-assisted result formats. The proportion of search pages containing AI elements is a changing product metric and is therefore omitted from this stable model-family article unless tied to a dated corporate report.

Qianfan platform

Baidu AI Cloud's Qianfan platform provides hosted model APIs, model development, retrieval and knowledge-base services, agents, workflow tools, evaluation, and deployment functions. As of 9 July 2026, its principal ERNIE endpoints included:

• ernie-5.1 for general text generation and agentic work;
• ernie-5.0 and ERNIE 5.0 thinking variants for text and native multimodal understanding;
• ERNIE 4.5 Turbo and ERNIE 4.5 Turbo VL variants;
• ERNIE X1.1 endpoints for deep-thinking workflows.

The catalogue's endpoint limits are service settings, not necessarily the maximum context supported by every underlying research checkpoint.

Coding and agents

Baidu integrates ERNIE into Baidu Comate, its software-development assistant, and into agent products for search, research, data analysis, and application construction. In its first-quarter 2026 reporting, Baidu highlighted products including DuMate, Miaoda 3.0, Famou Agent 2.0, and GenFlow 4.0.

The model alone is only one component of an agentic system. Reliability also depends on tool schemas, retrieval, permissions, execution sandboxes, state management, context compression, retry policies, and result verification.

Multimodal generation and understanding

ERNIE 4.5 VL supports image and video understanding, document question answering, OCR, chart and diagram interpretation, and visual reasoning. ERNIE 5.0 extends the hosted flagship line to native understanding and generation across text, image, video, and audio. ERNIE-Image provides a separately downloadable text-to-image generator.

A file type such as PDF is not itself an additional neural modality. PDF workflows usually parse text, render pages as images, or combine both before passing tokens to the model.

Domain applications

ERNIE-based systems have been used or marketed for education, finance, healthcare information, legal document processing, marketing, media production, digital humans, robotics, and Baidu's Apollo automotive ecosystem. Such deployment claims do not establish clinical, legal, or safety suitability. Domain systems require task-specific evaluation, access controls, auditability, and human oversight.

Open-weight releases and deployment

Baidu's downloadable ERNIE 4.5 checkpoints are licensed under Apache 2.0 and are available in PaddlePaddle and PyTorch-style formats. The associated stack includes ERNIEKit for fine-tuning and alignment and FastDeploy for serving. Official model cards also document support through frameworks such as Transformers, vLLM, and SGLang for selected checkpoints.

The largest standard ERNIE 4.5 checkpoints remain demanding. The official 300B-A47B text model card gives an example using 16 80-GB GPUs for BF16 serving and eight 80-GB GPUs for online FP8 quantization. Lower-bit community or Baidu research checkpoints can reduce the requirement, but memory, quality, and throughput depend on the exact quantization, context length, batch size, expert placement, and serving engine.

ERNIE-Image and ERNIE-Image-Turbo are also open-weight. As of 10 July 2026, Baidu's official Hugging Face organization listed ERNIE 4.5, ERNIE-Image, OCR, and other specialist models but did not list ERNIE 5.0 or 5.1 weights. Access to those flagships was therefore primarily through Baidu-operated products and Qianfan APIs.

Limitations and open problems

Resource requirements

MoE activation makes very large models computationally sparser, but it does not make their complete weight set disappear. Large ERNIE 4.5 and ERNIE 5.x systems require substantial storage, accelerator memory, high-bandwidth interconnects, expert load balancing, and fault-tolerant distributed serving. Long contexts add prefill computation and key-value-cache pressure.

ERNIE 5.1's smaller elastic subnetwork reduces the burden relative to ERNIE 5.0 according to Baidu, but the company has not disclosed enough deployment detail to calculate its exact memory footprint independently.

Hallucination and tool-use error

Knowledge-enhanced objectives, retrieval, verifiers, and adversarial training can reduce some error modes but do not eliminate hallucinations. ERNIE models may generate unsupported facts, false citations, incorrect calculations, or plausible but unsafe instructions. Agent systems introduce additional failure modes: choosing the wrong tool, sending malformed arguments, misreading a tool response, repeating actions, or propagating an early error through a long execution trace.

Model-generated reasoning text is not itself verification. High-impact claims should be checked against primary sources, and generated code should be tested in an isolated environment.

Dependence on knowledge graphs and synthetic data

Explicit knowledge objectives depend on entity linking, relation alignment, graph coverage, and the accuracy of the source knowledge base. Incorrect or incomplete triples can create misleading associations. Later generations depend heavily on synthetic instructions, model-generated solutions, verifiers, and simulated agent environments, raising questions about judge bias, reward hacking, benchmark contamination, and transfer to real production systems.

Transparency and reproducibility

Baidu has released model weights and tooling for ERNIE 4.5 and several specialist models, which enables independent inference, fine-tuning, and benchmark studies. It has not released the complete ERNIE 4.5 training dataset or the end-to-end training system, and the production ERNIE 5.0 and 5.1 checkpoints were not publicly listed as downloadable weights as of July 2026.

Consequently, third parties can reproduce some published inference experiments with open ERNIE 4.5 checkpoints but cannot exactly reproduce the original 4.5, 5.0, or 5.1 training runs from public artefacts alone.

Benchmark comparability

Vendor reports frequently change baseline versions, prompts, reasoning budgets, tool access, and judge models. A benchmark result may measure a complete harness rather than only the underlying model. Internal or newly introduced benchmarks provide useful evidence about intended workloads but generally offer less reproducibility than fixed public datasets with third-party runs.

Social bias

A 2025 PeerJ Computer Science study examined more than 30,000 generated descriptors for 240 Chinese social groups across 13 categories using Baidu Search, ERNIE, and Qwen. It found that Baidu Search and Qwen produced negative candidate words more frequently than the tested ERNIE service and identified a moderate prevalence of stereotypes in the language-model outputs.

The study evaluated specific commercial systems and prompts at a particular time. Its “negative word” and stereotype measures are proxies, not complete measures of social harm, and the results should not be generalized automatically to every ERNIE checkpoint or later release.

Medical safety

A 2025 simulated-patient study evaluated ERNIE Bot in 384 consultations involving unstable angina and asthma. The tested system achieved 77.3% diagnostic accuracy and 94.3% correct-drug-prescription accuracy, but requested unnecessary laboratory tests in 91.9% of trials and recommended unnecessary or inappropriate medications in 57.8%. The authors also reported different treatment intensity by age and household economic status.

These findings apply to the evaluated ERNIE Bot version, scenarios, and study design; they are not a measurement of the entire ERNIE model family. They nevertheless illustrate why a general chatbot should not be treated as an autonomous medical decision-maker.

Ethical and regulatory considerations

Public ERNIE services in China operate within the country's algorithm, data, deep-synthesis, and generative-AI regulatory framework. The Interim Measures for the Management of Generative Artificial Intelligence Services took effect on 15 August 2023 and set obligations concerning lawful data, security, personal information, and content governance for public-facing services.

China's Measures for Labeling Artificial Intelligence-Generated Synthetic Content, issued on 14 March 2025 and effective from 1 September 2025, require explicit user-visible labels and implicit metadata labels in applicable text, image, audio, video, and virtual-scene workflows.

A 2026 PNAS Nexus study compared China-origin and non-China-origin models on 145 politically sensitive questions in Chinese and English. The tested China-origin systems, including an ERNIE Bot service, showed substantially higher refusal rates and generally shorter and less accurate responses on those questions than the non-China models in the study.

That research concerns behavior under a defined political-question benchmark and does not evaluate the total capability, safety, or factuality of the models. It does, however, show that regulatory and product-level alignment can materially shape observable model behavior.

Deployers outside China may additionally be subject to privacy, copyright, consumer-protection, accessibility, product-safety, sector-specific, export-control, and AI-governance requirements in their own jurisdictions. Multimodal services can process personal images, voices, documents, or biometric information, making data minimization, consent, retention limits, and access control particularly important.

Research directions

Published ERNIE work indicates several continuing directions:

• Unified multimodal modelling. Improving joint text, image, video, and audio understanding and generation without sacrificing strong language performance.
• Elastic foundation models. Training one super-network from which subnetworks with different depth, expert capacity, and routing sparsity can be selected for different hardware and latency budgets.
• Agentic reinforcement learning. Scaling long-horizon environment interaction, tool use, verifier design, asynchronous rollout, and multi-capability policy integration.
• Efficient MoE serving. Reducing expert communication, memory duplication, load imbalance, and KV-cache pressure through disaggregated serving and quantization.
• Verifiable and grounded generation. Combining retrieval, structured knowledge, executable tools, and task-specific verifiers while reducing reward hacking and unsupported answers.
• Independent evaluation. Reproducing vendor results, measuring multilingual and multimodal safety, and testing agent systems in realistic environments rather than only static benchmarks.

The official materials do not provide a binding roadmap for a future ERNIE 5.2 or 6.0. Claims about unreleased version numbers or specifications would therefore be speculative as of 10 July 2026.

Comparison with other Chinese foundation-model families

ERNIE competes with Chinese model families including Qwen from Alibaba, GLM from Zhipu AI, DeepSeek, Hunyuan from Tencent, Doubao from ByteDance, and Kimi from Moonshot AI. Its historical differentiators include knowledge-oriented pre-training, close integration with Baidu Search and Baidu AI Cloud, and development on PaddlePaddle.

The competitive distinction has changed over time. Explicit knowledge-graph objectives were central to ERNIE 3.0 and Titan, whereas ERNIE 4.5 and 5.x are better characterized by multimodal MoE, elastic training, reasoning, and agents. Similarly, leaderboard positions are temporary snapshots and should not be used as enduring evidence that one family is categorically superior.

External links

• ERNIE official blog
• Qianfan model catalogue
• Baidu official organization on Hugging Face
• PaddlePaddle ERNIE repository
• FastDeploy repository

See also

• BERT
• Transformer (deep learning architecture)
• Mixture of experts
• Reinforcement learning from human feedback
• AI agent
• Qwen
• DeepSeek

Literature

• Sun, Y. et al. ERNIE: Enhanced Representation through Knowledge Integration. arXiv:1904.09223, 2019.
• Sun, Y. et al. ERNIE 2.0: A Continual Pre-training Framework for Language Understanding. AAAI, 2020.
• Sun, Y. et al. ERNIE 3.0: Large-scale Knowledge Enhanced Pre-training for Language Understanding and Generation. arXiv:2107.02137, 2021.
• Wang, S. et al. ERNIE 3.0 Titan: Exploring Larger-scale Knowledge Enhanced Pre-training for Language Understanding and Generation. arXiv:2112.12731, 2021.
• Feng, Z. et al. ERNIE-ViLG 2.0: Improving Text-to-Image Diffusion Model with Knowledge-Enhanced Mixture-of-Denoising-Experts. arXiv:2210.15257, 2022.
• Baidu ERNIE Team. ERNIE 4.5 Technical Report. 2025.
• ERNIE Team, Baidu. ERNIE 5.0 Technical Report. arXiv:2602.04705, 2026.
• Baidu ERNIE Team. ERNIE-Image Technical Report. arXiv:2605.25347, 2026.
• Liu, G.; Bono, C. A.; Pierri, F. Comparing diversity, negativity, and stereotypes in Chinese-language AI technologies. PeerJ Computer Science, 2025.
• Pan, J.; Xu, X. Political censorship in large language models originating from China. PNAS Nexus, 2026.
• Si, Y. et al. Quality, safety, and disparity of an AI chatbot in managing chronic diseases: simulated patient experiments. 2025.