Baidu open-sources ERNIE 4.5 Thinking, compact multimodal model taking on GPT 5
Baidu has released ERNIE 4.5 Thinking as open source, a compact multimodal AI model that, according to the first technical analyses, is set to compete with frontier models such as GPT 5 and Gemini. The model focuses on reasoning, code and multimodal input, and is positioned as an efficient building block for companies and developers who want to host or integrate AI functionality themselves.
According to the documentation and early benchmarks, ERNIE 4.5 Thinking combines strong language performance with support for image processing and advanced reasoning tasks. The release fits into a broader trend in which major providers are making not only heavyweight cloud models but also more compact variants available, so that developers can run AI closer to the application and in some cases even on-premises.
Baidu presents the model as a competitor to GPT 5 and Gemini 2.5 in applications such as code assistance, knowledge retrieval, analytical summaries and multimodal chat interfaces. Although independent, large-scale benchmarks are still limited, early measurements indicate that ERNIE 4.5 Thinking stays within the margin of the newest generations of closed-source models on a number of coding and reasoning tasks, while its hardware footprint is smaller. This makes the model interesting for organisations that want to combine their own infrastructure with high AI performance.
In practice, this step means that organisations have even more freedom of choice. Until recently, development teams were mainly dependent on proprietary APIs from a handful of US providers, but a landscape is now emerging in which open and semi-open models from different regions compete on latency, quality, privacy and cost efficiency. Integrations with existing open-source tooling, model hubs and orchestration platforms are expected to follow quickly in the coming weeks.
For companies with strict requirements around data residency and governance, the ability to run ERNIE 4.5 Thinking inside their own networks is particularly relevant. This offers greater control over auditability and logging, and allows fine-grained policy choices, for example over which data may or may not be sent to external clouds. At the same time, the trade-off between raw model capacity, maintenance burden and compliance remains complex, because organisations now have to assess a broader palette of AI building blocks.
The release of ERNIE 4.5 Thinking underlines how quickly the AI market is shifting towards a multi-model, multi-vendor reality. Instead of one central model for all tasks, more and more organisations are opting for a mix of heavyweight cloud models for the most complex generative tasks and compact models for latency-critical, privacy-sensitive or domain-specific workloads. Baidu's new model thus becomes one of the options in that strategy and is expected to surface rapidly in new developer frameworks, AI agents and industrial applications.
