ERNIE-4.5-21B-A3B-Thinking

What it is

ERNIE-4.5-21B-A3B-Thinking is an open-source large language model from Baidu built for deep thinking: instead of answering immediately, it works through a problem in a longer chain of reasoning before producing a result. It is a mixture-of-experts model derived from ERNIE-4.5-21B-A3B, with 21 billion total parameters and 3 billion active parameters^[2]. Its context window is 128K, or 131,072 tokens^[2].

"Mixture of experts" means the model contains many separate sub-networks, called experts, and routes each piece of input to only a few of them. Because only the selected experts are used for any given token, a model can hold far more parameters than it actually computes with, which keeps inference cheaper than the total parameter count suggests. The "21B-A3B" in the name records exactly that split: 21 billion total, 3 billion active.

Where it came from

Baidu open-sourced the model at WAVE SUMMIT 2025 on 9 September 2025, the same day the company presented ERNIE X1.1^[1]. It followed the 30 June open-sourcing of ten ERNIE 4.5 models, which established the family it belongs to^[1]. The weights were released under the Apache License 2.0, a license that permits commercial use, and are distributed on Hugging Face and Baidu AI Studio^[2][1].

What it is built for

The model targets complex reasoning tasks, including logic, mathematics, science, code generation and content creation, and it supports tool use through function calling^[1][2]. Its model card recommends it specifically for highly complex reasoning tasks rather than as a general-purpose chat model^[2]. Compared with the earlier version, the model card cites a longer thinking length and improved understanding of long contexts up to 128K^[2].

Inside the model

The model card gives the expert layout: 64 text experts in total, of which 6 are activated for a given token, plus 2 shared experts^[2]. The shared experts are used alongside whichever routed experts are selected, a common way to give mixture-of-experts models a stable baseline of computation.

Running it

Because it is released openly, the model can be run locally or on your own infrastructure, and the model card is explicit about the software versions it expects. It lists FastDeploy 2.2 as required, vLLM 0.10.2 or later but not 0.11.0, and Transformers 4.54.0 or newer^[2]. It also names SGLang and Docker Model Runner among the local applications that can run the model^[2]. For teams that want to adapt it rather than just serve it, Baidu's ERNIEKit added supervised fine-tuning and function-call training for the model in September 2025^[3].

Where it sits

The model is a small, openly licensed member of the ERNIE 4.5 generation, specialized for reasoning. Its design reflects a broader pattern in open model releases: take a compact mixture-of-experts base, add a longer thinking process, and publish the weights under a permissive license so others can run, fine-tune and build on it. Its 3 billion active parameters place it in the range of models that can be served without a large cluster, while the 128K context and function calling make it usable for multi-step tasks that involve long documents and external tools.