# Baidu opened ERNIE 4.5 as a model family, making artifact and license selection part of deployment Baidu opened ten ERNIE 4.5 variants on June 30, 2025, with Apache 2.0 terms, public artifacts and development tooling. Canonical: https://rangoon.ai/insights/archive-baidu-ernie-4-5-open-model-family-june-2025/ Author: Rangoon Editorial (https://rangoon.ai/insights/editorial/) Published: 2026-10-24 Updated: 2026-10-24 Event date: 2025-06-30 Tags: [open source](https://rangoon.ai/insights/tags/open-source/), [multimodal AI](https://rangoon.ai/insights/tags/multimodal-ai/), [mixture of experts](https://rangoon.ai/insights/tags/mixture-of-experts/), [Apache 2.0](https://rangoon.ai/insights/tags/apache-2-0/) Source note: Historical reporting: Baidu announced an open-source ERNIE 4.5 family on June 30, 2025. Its launch post names ten variants, including mixture-of-experts and dense models, says the models were publicly accessible under Apache 2.0, and links public code and model destinations. The June technical report provides architecture and evaluation context. Vendor benchmark and efficiency statements remain Baidu-reported; public artifacts and an Apache license do not prove local compatibility, current availability, security, or a deployment outcome. Image: Versioned multimodal model-release record. Original Rangoon explanatory diagram: Source boundary: dated primary release reporting; original editorial analysis. No vendor interface, photograph, measurement result, or endorsement implied. ## Key takeaways - The June event was an open release of a family of named artifacts, not one interchangeable ERNIE 4.5 configuration. - Apache 2.0 establishes a stated license for the release, while model size, modality, runtime and derivative packaging remain deployment decisions. - Baidu-reported benchmarks and training-efficiency figures describe its methods and tests; they do not predict a separate workload. ## Baidu dated the public release to June 30 and described ten distinct variants [source 1](https://ernie.baidu.com/blog/posts/ernie4.5/) [source 2](https://ernie.baidu.com/blog/publication/ERNIE_Technical_Report.pdf) On June 30, 2025, Baidu published an English ERNIE Blog post titled “Announcing the Open Source Release of the ERNIE 4.5 Model Family.” The post says the family comprised ten variants. It describes mixture-of-experts models with either 47B or 3B active parameters, a largest listed model with 424B total parameters, and a 0.3B dense model. Those identifiers matter because an ERNIE 4.5 family label alone does not say which checkpoint, mode, parameter footprint or input type an operator selected. The dated record supports an open family release with these stated categories. It does not establish that every variant fits the same accelerator, exposes the same interface, or is suitable for the same text or image task. ## The release connected public artifacts to Apache 2.0 without turning “open” into a deployment conclusion [source 1](https://ernie.baidu.com/blog/posts/ernie4.5/) Baidu’s post says the models were publicly accessible under Apache 2.0 and that it also open sourced development toolkits for ERNIE 4.5. The same page links GitHub, Hugging Face and its own tooling. It later says the models are available in PyTorch-compatible formats, while its examples also use the PaddlePaddle ecosystem. Those are useful distribution and tooling facts, but they are not a promise that every hosted endpoint, conversion, quantization or third-party package has identical behavior. A license should be reviewed against the exact artifact and planned use, and a locally converted checkpoint should be recorded as a separate supply-chain object. The release makes inspection and adaptation possible; it does not supply a deployment’s access control, retention rules, vulnerability handling or acceptance test. ## Heterogeneous MoE and modality-specific post-training described a design, not an automatic multimodal result [source 1](https://ernie.baidu.com/blog/posts/ernie4.5/) [source 2](https://ernie.baidu.com/blog/publication/ERNIE_Technical_Report.pdf) The launch post characterizes ERNIE 4.5 as a multimodal family with a heterogeneous MoE structure, including shared and modality-specific parameters. It also distinguishes language-model variants from vision-language variants and says the latter support thinking and non-thinking modes. These are model-family design statements. They do not show that a particular document, image or mixed input will be understood correctly, nor that changing a thinking mode preserves a caller’s schema, latency budget or safety policy. An evaluation needs an exact checkpoint, tokenizer or processor revision, image preprocessing route, mode selection, decoding settings and permitted fixtures. The June post also reports benchmark outcomes and training figures. Those are Baidu’s results under its chosen methods, not independent evidence that a separate system will preserve accuracy or inference efficiency. ## A model-family release made provenance more important for operators [source 1](https://ernie.baidu.com/blog/posts/ernie4.5/) [source 2](https://ernie.baidu.com/blog/publication/ERNIE_Technical_Report.pdf) For an operator, the practical result of a ten-variant release is a selection record rather than a generic model claim. The record should name the artifact, source revision, license review, runtime, quantization, hardware profile, input contract and observed test result. Baidu’s examples mention low-bit inference, context caching, speculative decoding and an OpenAI-compatible API through a deployment toolkit. Those are possible implementation paths described by the publisher, not evidence that every implementation exposes the same behavior or is safe for a connected workflow. The historical conclusion is narrow: Baidu opened a named ERNIE 4.5 family and associated tooling on June 30. It does not establish present service terms, enterprise compliance, a universal cost result or authorization for a model output to act beyond its application boundary. ## Sources 1. [ERNIE Blog: Announcing the Open Source Release of the ERNIE 4.5 Model Family](https://ernie.baidu.com/blog/posts/ernie4.5/) 2. [Baidu ERNIE Team: ERNIE 4.5 Technical Report](https://ernie.baidu.com/blog/publication/ERNIE_Technical_Report.pdf) ## Related reading - [Tencent released Hunyuan-A13B under a territory-limited community license, separating open weights from universal permission](https://rangoon.ai/insights/archive-tencent-hunyuan-a13b-open-hybrid-reasoning-june-2025/) - [PaddleOCR 3.0 put OCR, document parsing and extraction into one versioned release](https://rangoon.ai/insights/archive-paddleocr-3-document-parsing-may-2025/) - [Microsoft’s VibeVoice release put long-form multi-speaker TTS into an open research framework, then exposed its misuse boundary](https://rangoon.ai/insights/archive-microsoft-vibevoice-open-source-tts-august-2025/) ## More information - [Documentation index](https://rangoon.ai/llms.txt) - [AI access and policies](https://rangoon.ai/ai/) - [Source repository](https://github.com/hypler-dev/rangoon)