Parameter	Value
direction_index	per layer
attn.o_proj.max_weight	1.60
attn.o_proj.max_weight_position	26.32
attn.o_proj.min_weight	1.31
attn.o_proj.min_weight_distance	4.71
mlp.down_proj.max_weight	0.88
mlp.down_proj.max_weight_position	19.16
mlp.down_proj.min_weight	0.01
mlp.down_proj.min_weight_distance	2.78

Performance

Metric	This model	Original model (baidu/ERNIE-4.5-21B-A3B-Base-PT)
KL divergence	0.3169	0 (by definition)
Refusals	16/100	94/100

Github

Discord

ERNIE-4.5-21B-A3B-Base

[!NOTE] Note: "-Paddle" models use PaddlePaddle weights, while "-PT" models use Transformer-style PyTorch weights.

[!NOTE] Note: The Base model only supports text completion. For evaluation, use the completion API (not chat_completion) in vLLM/FastDeploy.

ERNIE 4.5 Highlights

The advanced capabilities of the ERNIE 4.5 models, particularly the MoE-based A47B and A3B series, are underpinned by several key technical innovations:

Multimodal Heterogeneous MoE Pre-Training: Our models are jointly trained on both textual and visual modalities to better capture the nuances of multimodal information and improve performance on tasks involving text understanding and generation, image understanding, and cross-modal reasoning. To achieve this without one modality hindering the learning of another, we designed a heterogeneous MoE structure, incorporated modality-isolated routing, and employed router orthogonal loss and multimodal token-balanced loss. These architectural choices ensure that both modalities are effectively represented, allowing for mutual reinforcement during training.
Scaling-Efficient Infrastructure: We propose a novel heterogeneous hybrid parallelism and hierarchical load balancing strategy for efficient training of ERNIE 4.5 models. By using intra-node expert parallelism, memory-efficient pipeline scheduling, FP8 mixed-precision training and finegrained recomputation methods, we achieve remarkable pre-training throughput. For inference, we propose multi-expert parallel collaboration method and convolutional code quantization algorithm to achieve 4-bit/2-bit lossless quantization. Furthermore, we introduce PD disaggregation with dynamic role switching for effective resource utilization to enhance inference performance for ERNIE 4.5 MoE models. Built on PaddlePaddle, ERNIE 4.5 delivers high-performance inference across a wide range of hardware platforms.
Modality-Specific Post-Training: To meet the diverse requirements of real-world applications, we fine-tuned variants of the pre-trained model for specific modalities. Our LLMs are optimized for general-purpose language understanding and generation. The VLMs focuses on visuallanguage understanding and supports both thinking and non-thinking modes. Each model employed a combination of Supervised Fine-tuning (SFT), Direct Preference Optimization (DPO) or a modified reinforcement learning method named Unified Preference Optimization (UPO) for post-training.

To ensure the stability of multimodal joint training, we adopt a staged training strategy. In the first and second stage, we train only the text-related parameters, enabling the model to develop strong fundamental language understanding as well as long-text processing capabilities. The final multimodal stage extends capabilities to images and videos by introducing additional parameters including a ViT for image feature extraction, an adapter for feature transformation, and visual experts for multimodal understanding. At this stage, text and visual modalities mutually enhance each other. After pretraining trillions tokens, we extracted the text-related parameters and finally obtained ERNIE-4.5-21B-A3B-Base.

Model Overview

ERNIE-4.5-21B-A3B-Base is a text MoE Base model, with 21B total parameters and 3B activated parameters for each token. The following are the model configuration details:

Key	Value
Modality	Text
Training Stage	Pretraining
Params(Total / Activated)	21B / 3B
Layers	28
Heads(Q/KV)	20 / 4
Text Experts(Total / Activated)	64 / 6
Vision Experts(Total / Activated)	64 / 6
Shared Experts	2
Context Length	131072

Quickstart

Using `transformers` library

Note: You'll need the transformers library (version 4.54.0 or newer) installed to use this model.

The following contains a code snippet illustrating how to use the model generate content based on given inputs.

python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "baidu/ERNIE-4.5-21B-A3B-Base-PT"

# load the tokenizer and the model
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    device_map="auto",
    torch_dtype=torch.bfloat16,
)

prompt = "Large language model is"
model_inputs = tokenizer([prompt], add_special_tokens=False, return_tensors="pt").to(model.device)

generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=1024
)
result = tokenizer.decode(generated_ids[0].tolist(), skip_special_tokens=True)
print("result:", result)

vLLM inference

vllm>=0.10.2 (excluding 0.11.0)

vllm serve baidu/ERNIE-4.5-21B-A3B-Base-PT

License

Citation

If you find ERNIE 4.5 useful or wish to use it in your projects, please kindly cite our technical report:

bibtex

@misc{ernie2025technicalreport,
      title={ERNIE 4.5 Technical Report},
      author={Baidu ERNIE Team},
      year={2025},
      eprint={},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={}
}

trohrbaugh/ERNIE-4.5-21B-A3B-Base-PT-heretic

This is a decensored version of baidu/ERNIE-4.5-21B-A3B-Base-PT, made using Heretic v1.2.0

Abliteration parameters