ERNIE-4.5-21B-A3B-Base
Active Parameters
21B
Context Length
131K
Modality
Text
Architecture
Mixture of Experts
KV Cache Rate
56 KB/tok
License
Apache 2.0
Release Date
30 Jun 2025
Knowledge Cutoff
Dec 2024
API Pricing (per 1M)
Self-hosted only
Hardware & System Requirements
Precision
Context
Workload Memory: 14.11 GB
INT4 · 8K context
Base Weights: 11.18 GB
KV Cache: 0.44 GB
Runtime & Graphs: 2.50 GB
Minimum Required Hardware
1x NVIDIA L4 24GB
14.1 / 21.6 GB usable
1x RTX 4060 Ti 16GB
14.1 / 14.4 GB usable
Apple M4 24GB
14.1 / 18.0 GB usable
Determine requirements for your specific workload:
Open in VRAM Calculator
Runtime requirements
Checkpoint format
FP16
CUDA
>= 11.8
PyTorch
>= 2.1
Transformers
>= 4.54.0.dev0
| Architecture | Precision | Attention kernel |
|---|---|---|
Blackwell, Hopper | Native FP16 | FlashAttention-3 |
Ada Lovelace, Ampere | Native FP16 | FlashAttention-2 |
Architecture Diagram
Evaluation Benchmarks
No evaluation benchmarks for ERNIE-4.5-21B-A3B-Base available.
Rankings
Overall Rank
-
Coding Rank
-
About ERNIE-4.5-21B-A3B-Base
ERNIE-4.5-21B-Base is an open-weights Mixture-of-Experts text model by Baidu activating 3B parameters per token for efficient NLP applications. It utilizes modality-isolated routing and progressive RoPE scaling across a 131K context window.
Specifications
Attention
Attention Structure
Grouped-Query Attention
Attention Heads
20
Key-Value Heads
4
Attention Head Dimension
128
Position Embedding
Absolute Position Embedding
RoPE Theta
500,000
Normalization
RMS Normalization
Activation Function
SwigLU
Dimensions
Base Precision
FP16
Hidden Dimension Size
2,560
Number of Layers
28
FFN Intermediate Size (Dense)
1,536
Tokenizer
Vocabulary Size
103,424
Mixture of Experts
Total Expert Parameters
3.0B
Number of Experts
64
Active Experts
6
Shared Experts
2
FFN Intermediate Size (per Expert)
1,536
Dense Layers Before MoE
1
About ERNIE 4.5
The Baidu ERNIE 4.5 family consists of ten large-scale multimodal models. They utilize a heterogeneous Mixture-of-Experts (MoE) architecture, which enables parameter sharing across modalities while also employing dedicated parameters for specific modalities, supporting efficient language and multimodal processing.