Baidu open-sources ERNIE-4.5-21B-A3B-Thinking

Background

Baidu has been releasing open versions of its ERNIE 4.5 family. On 30 June 2025, the company open-sourced ten ERNIE 4.5 models, making them available for outside developers to download and run.^[1] The September release is a deep-thinking model. In practice, a deep-thinking model works through intermediate reasoning steps before producing a final answer, rather than emitting a response immediately. That style of processing is useful when a task benefits from being broken into smaller steps, as with logic puzzles, mathematics and code.

The release

On 9 September 2025, at WAVE SUMMIT 2025, Baidu open-sourced ERNIE-4.5-21B-A3B-Thinking.^[1] The model was released as open weights, so developers can obtain the parameters rather than only call a hosted service. Baidu targets it at logic, mathematics, code generation and tool use, and the model supports function calling.^[1][2] Function calling lets a model emit a structured request for an external tool or service; the surrounding application then executes that request and can pass the result back to the model.

Architecture and context

ERNIE-4.5-21B-A3B-Thinking is a mixture-of-experts (MoE) model with 21 billion total parameters and 3 billion active parameters.^[1][2] In a mixture-of-experts model, each input token is routed to a subset of the network's expert modules instead of being passed through every parameter. The total parameter count is therefore larger than the number of parameters used for any single token. This matters for serving: memory requirements scale with the total size, while the computation per token is closer to the active size. Baidu lists a 128K context window,^[1][2] and the Hugging Face configuration gives a context length of 131,072 tokens.^[2] A larger context window allows more text—code, documents or conversation history—to be processed in one pass.

Reported capabilities

The Hugging Face model card states that the model shows significantly improved performance in logical reasoning, mathematics, science, coding and text generation.^[2] These areas match the tasks Baidu highlights for the release. Mathematics and coding are multi-step tasks where intermediate work affects the final result, while tool use requires the model to decide when an external function is needed and how the request should be formed.

Availability and serving

The model is available under the Apache 2.0 license on Hugging Face and Baidu AI Studio.^[1][2] Apache 2.0 is a widely used permissive open-source license. The model card gives serving instructions for Transformers, vLLM, SGLang, Docker Model Runner and FastDeploy.^[2] These instructions give developers several supported paths for loading and running the model.

Place in the ERNIE line

The release followed the 30 June 2025 open-sourcing of ten ERNIE 4.5 models.^[1] Baidu thus added an open-weight reasoning model to a family that already had openly released models. For developers, the significance is concrete: a model aimed at multi-step reasoning, mathematics, coding and tool use can be downloaded, inspected and served on their own infrastructure under a standard open-source license.