ERNIE-4.5-21B-A3B-Thinking

ERNIE-4.5-21B-A3B-Thinking

Reasoning-focused ERNIE 4.5 mixture-of-experts text model with 21B total and 3B active parameters, 128K context and tool calling, in English and Chinese. For developers who need deep reasoning from a small active size.

Made by baidu on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: This is Baidu's thinking-tuned release of ERNIE 4.5 21B-A3B, built to reason at length over hard problems. It is a mixture-of-experts model with about 21B parameters in total, of which 3B are active for each token. Baidu says reasoning, tool use and long-context understanding were improved over the earlier A3B release, and this version thinks for longer.

Who it suits: It suits developers and researchers who want a reasoning model that fits on one 80 GB GPU and who can run a model server.

Industry | Who uses it | How they use it | Example
Education | Teachers and students | Work through math and science problems step by step | Checking a physics derivation
Software | Developers | Use it for coding help and for calling functions | Asking it to pick a function and fill in its arguments
Research | Analysts | Reason over long documents within the 128K context | Summarizing a long report with reasoning
IT operations | Platform engineers | Serve a mid-size reasoning model behind an OpenAI-style API | Running it with FastDeploy

If your tasks are simple chat, you can skip it, since the extra thinking length adds waiting time.

Where to use it: You run it as a server and call it from your own apps.

• FastDeploy, which the card shows with a full launch command
• Other tools that load Transformer-style weights, such as vLLM and transformers, which the card expects to work
• Windows, macOS or Linux with Python, to download the files

Use case | Who | What they get back
Hard reasoning questions | Researchers | A long thinking trace, then an answer
Function calling | Developers | A tool call chosen from the functions you describe
Long inputs | Analysts | Answers over up to 131072 tokens of context

How to set it up: The card says serving with FastDeploy needs one 80 GB GPU and FastDeploy version 2.2. The weights are about 43.7 GB in BF16.

• One GPU with 80 GB of memory
• FastDeploy version 2.2, plus Python
• Disk space for the model files, about 43.7 GB

1. Install FastDeploy 2.2 by its own instructions.
2. Download the model, or let the server fetch it by name.
3. Start the API server with the command below.

pip install -U huggingface_hub
hf download baidu/ERNIE-4.5-21B-A3B-Thinking

How to use it: Start the server, then send chat requests to its OpenAI-style endpoint. The same launch also turns on the reasoning parser and the tool call parser.

Tool | What it does
fastdeploy.entrypoints.openai.api_server | Serves the model through an OpenAI-style API
--reasoning-parser ernie_x1 | Separates the model's thinking from its answer
--tool-call-parser ernie_x1 | Reads function calls out of the output
--max-model-len 131072 | Sets the full context length

python -m fastdeploy.entrypoints.openai.api_server \
       --model baidu/ERNIE-4.5-21B-A3B-Thinking \
       --port 8180 \
       --metrics-port 8181 \
       --engine-worker-queue-port 8182 \
       --load-choices "default_v1" \
       --tensor-parallel-size 1 \
       --max-model-len 131072 \
       --reasoning-parser ernie_x1 \
       --tool-call-parser ernie_x1 \
       --max-num-seqs 32

Example: You send a chat request that asks about today's weather in Beijing and describes a get_weather function with a location and a unit. The model supports function calls, so it can reply with a call to that function instead of guessing.

For newcomers and experts: A thinking model writes out its reasoning first, then gives the answer. Start with one hard question, such as a short math problem, and read the reasoning it returns.

For experts, it is a text-only mixture-of-experts model after post-training. It has 28 layers, 20 query heads with 4 key-value heads, and 64 text experts with 6 active per token, plus 2 shared experts. The trade-off is a longer thinking length, which costs latency.

Your prompt goes to the server
The router picks 6 of 64 experts
The model writes its thinking first
Then it writes the answer or a tool call
The parsers split thinking from output

Good to know: Baidu released Transformer-style weights so that PyTorch and PaddlePaddle tools can load them. Others have published 24 quantized and 4 fine-tuned versions of this model.

Title: Model configuration from the card
Setting | Value
Total and active parameters | 21B and 3B
Layers | 28
Text experts, total and active | 64 and 6
Shared experts | 2
Context length | 131072

• License: Apache 2.0
• Cost to run: none named, beyond one 80 GB GPU
• Data it reaches: the prompts you send to your own server, and the model files you download
• Not for: simple, quick chat, since the card recommends it for highly complex reasoning

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 21.8B
• Weights: about 43.7 GB in BF16; about 10.9 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 64,420
• Likes: 796
• License: apache-2.0
• Languages: en, zh
• Quantized versions: 24
• Fine-tunes: 4
• Tags: transformers, safetensors, ernie4_5_moe, text-generation, ernie4.5, conversational

Links

• Model: https://huggingface.co/baidu/ERNIE-4.5-21B-A3B-Thinking
• Files: https://huggingface.co/baidu/ERNIE-4.5-21B-A3B-Thinking/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download baidu/ERNIE-4.5-21B-A3B-Thinking

Then follow the usage in its model card. Nothing in this kit runs by itself.