424B- Context
- 131Ktokens
- Released
- 1y ago
- Size
- 21.9Bparameters
Baidu open-weight 21B mixture-of-experts language model with 3B active parameters.
How it compares
Additional source evidence
- Source model
- baidu/ernie-4.5-21b-a3b
- Source family
- ernie
- Last updated
- 2025-06-30
- Model context
- 131K
- Model output limit
- 8K
- Open weights
- Yes
- File attachments
- No
- Reasoning
- No
- Tool calling
- Yes
- Temperature control
- Yes
Downloadable weights
Complete source record (JSON)
{
"id": "baidu/ernie-4.5-21b-a3b",
"name": "ERNIE 4.5 21B A3B",
"description": "Open ERNIE text MoE with 21B total and 3B active parameters",
"family": "ernie",
"last_updated": "2025-06-30",
"attachment": false,
"reasoning": false,
"tool_call": true,
"temperature": true,
"modalities": {
"input": [
"text"
],
"output": [
"text"
]
},
"limit": {
"context": 131072,
"output": 8000
},
"release_date": "2025-06-30",
"open_weights": true,
"license": "Apache-2.0",
"weights": [
{
"label": "Hugging Face",
"url": "https://huggingface.co/baidu/ERNIE-4.5-21B-A3B-PT"
}
]
}Frequently asked questions
Running locally
ERNIE 4.5 21B A3B needs an estimated 15.8 GB of VRAM at Q4, 27.3 GB at Q8 and 51 GB at FP16, based on its 21.9B parameters. These figures cover the weights plus about 15% and 0.5 GB for a short context, so they are estimates.
Yes on a 24 GB or 16 GB GPU, which fits the model at Q4 (estimated 15.8 GB). A 12 GB or 8 GB GPU does not fit it at Q4, Q8 or FP16, though these are estimates.
Yes. At Q4 it needs an estimated 15.8 GB, assuming macOS lets the GPU use about 75% of unified memory, so the smallest common Mac memory size that fits is 24 GB.
Privacy and data
Yes, ERNIE 4.5 21B A3B has open weights, so you can self-host it for full control of your data.
Capabilities and specs
The context window is 131K tokens (131,072), and the maximum output is 8K tokens (8,000).
No, it accepts text only. Its input modalities are limited to text.
Yes, tool calling (function calling) is supported.
No, ERNIE 4.5 21B A3B does not support a reasoning or thinking mode.
Quality and benchmarks
The closest alternatives, matched on model kind and release date with no comparable price, are DiffuCoder 7B by Apple, FlexOlmo 7x7B by Ai2 and Kimi K2 Instruct by Moonshot AI.
DiffuCoder 7BERNIE 4.5 21B A3B vs DiffuCoder 7BFlexOlmo 7x7BERNIE 4.5 21B A3B vs FlexOlmo 7x7BKimi K2 InstructERNIE 4.5 21B A3B vs Kimi K2 Instruct
Licensing and access
ERNE 4.5 21B A3B has open weights, released under the Apache 2.0 license.
Yes. The model is under the Apache 2.0 license, a permissive license that allows commercial use.
The weights are available on the Hugging Face repository baidu/ERNIE-4.5-21B-A3B-PT.
Basics
ERNIE 4.5 21B A3B was made by Baidu and released on June 30, 2025.
Yes, the model has open weights.
Generated from Vertical Tabs data, updated November 3, 2026.
Discussion
No comments yet
Be the first to share your thoughts.