Baidu releases open source ERNIE vision model, claiming benchmark wins over GPT-5 and Gemini

A 28 billion parameter model with a smaller working core

Baidu has released ERNIE 4.5 VL 28B A3B Thinking, an open source artificial intelligence model that processes images, video, documents and text. As of November 12, 2025, the company says the model matches or beats versions of OpenAI GPT-5 High and Google Gemini 2.5 Pro on selected vision benchmarks, while running inference on a single graphics processing unit with 80 GB of memory.

The model contains 28 billion parameters, the numerical settings learned during training, but activates about 3 billion for each token it processes. A token is a small unit of information handled by the model. Baidu uses a Mixture of Experts architecture, which directs each input to selected specialist networks instead of making the entire system work on every step.

Baidu has released the model under Apache 2.0, allowing unrestricted commercial use. Its combination of open licensing, visual reasoning and a relatively small active processing core gives developers another option for building systems that read documents, interpret diagrams and examine video without relying entirely on closed AI services.

The central qualification is that independent laboratories have not yet verified the performance charts. Baidu's benchmark claims are the company's own account of how the model compares with leading competitors. For businesses, the next practical step is testing it against the documents, images and operating conditions their applications actually encounter.

The release covers six areas: visual reasoning, science and mathematics problem solving, visual grounding, tool integration, video understanding and dynamic image analysis. Together, they describe a system intended to examine visual evidence, identify relevant details and work through questions that require several steps.

What Baidu says the benchmarks show

Baidu's documentation presents ERNIE 4.5 VL 28B A3B Thinking as matching or surpassing GPT-5 High and Gemini 2.5 Pro on a group of tests focused on vision. The tasks include answering questions about charts, understanding documents and solving visual problems through several stages of reasoning.

The company also claims inference speeds two to three times faster than comparable models with similar accuracy. Inference means using a trained model to produce an answer, rather than training it. Faster inference can shorten the wait for a response and increase the number of tasks a system can process.

Baidu's claims put the emphasis on visual workloads: reading information from a page, understanding the relationship between elements in a diagram, or interpreting a chart. These are important tasks for enterprises whose information arrives in scans, photographs and other formats that cannot be handled as plain text alone.

Benchmark results depend on the test. Some evaluations use clean images with carefully prepared crops; others include questions that overlap with material encountered during training. Scores can also respond to prompt formatting, pattern recall or narrow techniques that work especially well on a particular set of questions.

Companies considering the model therefore need to measure its accuracy on their own tasks. A document extraction system must recover the right fields, while a factory inspection system must identify the right object under the lighting and camera conditions of the production line.

How selective processing reduces the workload

ERNIE's Mixture of Experts design places specialist subnetworks behind a router. For each token or image patch, the router chooses a small subset of experts to perform the computation. Most of the model's parameters remain inactive during that operation, while the larger collection of experts provides capacity across different tasks and processing steps.

About 3 billion active parameters out of 28 billion total means roughly 11% of the model's parameters are active per token. That explains the difference between the two numbers in its name: the total size describes the pool of learned settings, while the active size describes the portion used for a particular step.

Baidu says selective activation reduces computation and the memory footprint, allowing inference on one GPU with 80 GB of memory. The company presents this as a way to retain much of a larger model's capacity while doing less work at each step. A dense model, by comparison, uses its full network rather than selecting experts.

The architecture also adds engineering complexity. The router must be fast and distribute work effectively. Training needs to keep a few experts from taking most of the workload, and deployment teams must watch for routing hotspots that can cause delays. Useful throughput depends on how well those moving parts work together.

Thinking with Images changes where the model looks

Baidu calls the model's central feature Thinking with Images. Instead of taking a single fixed view of an image and immediately producing an answer, the system can plan where to look, zoom into a region, inspect details and return to the wider scene.

The process starts by breaking a question into steps. The model identifies an area worth examining, crops it at higher resolution, reads text or examines geometry within that area, and updates its plan. It repeats the process until it has gathered enough information to answer.

A dense technical diagram illustrates the value of this approach. The whole page shows how components relate to one another, but a small label or symbol may supply the detail needed to solve the problem. Moving between those two scales allows the model to combine local clues with the broader arrangement.

Baidu also describes the ability to zoom out as well as in. That matters when a close inspection supplies only part of the answer. A request to locate the blue connector under the second bolt, for example, requires both recognition of a small object and an understanding of its position relative to other objects.

Grounding connects words to places in an image

Visual grounding means connecting a description to a precise object or region in an image. ERNIE's capabilities include mapping words back to pixel locations and indicating the relevant area with a box or a mask. A mask marks the shape or area of an object more closely than a rectangular box.

Manufacturing is one possible application. Two parts may look almost identical until a system examines a small feature. An inspection tool that identifies the relevant region can help workers understand where a suspected defect is located, rather than providing only a general description.

Engineering drawings present a similar challenge because a few pixels of notation can change a symbol's meaning. Medical imaging also depends on small details, where tiny artifacts can affect a diagnosis. These examples explain why the ability to inspect fine features is useful in demanding visual tasks.

The combination of broad context, close inspection and explanations of the steps taken can give operators a clearer audit trail and a way to correct mistakes. Baidu also highlights tighter instruction following for image tasks, helping connect the user's request with the area the system examines.

Documents, scientific problems and video

ERNIE's six capability areas cover work that extends beyond recognizing objects. The model can read and summarize invoices or financial statements, analyze charts and tables, and reason through circuit diagrams and geometry problems. STEM tasks involve science, technology, engineering and mathematics, where the answer may depend on both visual details and calculations.

Tool integration includes image search, allowing the model to retrieve reference pictures while working on a task. Combined with grounding and reasoning, that opens the possibility of workflows in which the system gathers information, examines it and takes several steps toward an answer.

Video understanding adds the need to follow events over time. Baidu describes temporal awareness, meaning the model can track what happens across video segments rather than treating every frame as an unrelated picture. The sequence of actions can matter as much as the contents of any one frame.

Potential applications include reading claims in insurance queues, flagging unusual activity in warehouse operations and helping technicians troubleshoot equipment using annotated photographs and short videos. These examples combine several abilities: interpreting a document or scene, locating relevant evidence and reasoning about what to do next.

The training recipe and tools for developers

The ERNIE 4.5 family combines supervised learning with reinforcement learning on multimodal tasks. Supervised learning trains a model using examples with target answers. Reinforcement learning uses rewards to encourage responses or actions that meet the training objective. Multimodal training brings different kinds of input, such as text and images, into that process.

The documentation names GSPO and IcePop among strategies used to stabilize the Mixture of Experts router and improve reward shaping. Reward shaping is the design of feedback that guides the model toward useful behavior. Dynamic difficulty sampling changes the difficulty of training examples as the model improves, introducing harder problems over time.

Baidu presents this training approach as supporting reasoning through several steps over pictures, charts and video, rather than making a single immediate guess. The company also says the model underwent multimodal reinforcement learning and grounding alignment, training intended to strengthen the connection between visual evidence and responses.

The wider ERNIE 4.5 family includes models for text alone and vision language variants. It ranges from very small dense systems to Mixture of Experts models with hundreds of billions of total parameters. Baidu says its architecture shares parameters across modalities so that text and vision components can benefit from each other while maintaining language quality.

Training and inference use Baidu's PaddlePaddle platform. The team reports high hardware utilization at large scale, meaning the computing equipment spends a large share of its time doing useful processing.

Developers can use ERNIEKit for training and fine tuning, including preference optimization and quantization. Fine tuning adapts a model to a task, preference optimization trains it toward preferred responses, and quantization represents numerical settings more compactly to support deployment with fewer resources.

The model also comes with integration kits for Transformers, vLLM and FastDeploy. Transformers and vLLM provide familiar routes for developers to work with and serve models, while FastDeploy supports production inference. Baidu provides an official model page on Hugging Face, with code and recipes in the Baidu code repository.

What businesses can test with one GPU

Open licensing and inference on a single GPU make ERNIE attractive to organizations seeking visual AI on their own premises. Such deployments let teams process documents and images without sending them to a cloud provider, an important consideration when security policies restrict where information can go.

Document intensive businesses could test intake systems for contracts, invoices and insurance claims. Manufacturers could build inspection stations that combine visual analysis with language guidance for workers. Retail and logistics teams could locate products or verify labels in warehouses, while customer service teams could use charts and screenshots to improve answers.

The hardware requirement matters to cost. Many organizations cannot reserve clusters of GPUs in the H100 class for one application. Running a model on a single card with 80 GB of memory lowers the entry barrier for pilots and early deployments, even though that remains substantial equipment.

Grounding, tool use and video understanding also suit workflows carried out by AI agents. In these workflows, a model reads, searches, clicks and reasons through several actions to complete a task. That fits the move toward agents that operate ordinary computer interfaces rather than relying entirely on custom application programming interfaces, or APIs.

A useful pilot needs representative inputs. Businesses should test their own documents, lighting conditions, camera angles and security requirements before committing to production. The task is to find out how reliably the system handles the material that matters to the organization.

Why real operating conditions need their own tests

Real documents and images introduce glare, motion blur, unusual fonts, blocked objects and labels in multiple languages. A model's performance can change when it encounters inputs unlike those used in training. The same system may handle a neatly scanned page well and struggle with a photograph of that page taken at an angle.

For document processing, teams can assemble a few hundred representative files and measure exact field extraction accuracy against answers checked by people. This tests whether the system recovers the required information correctly, rather than whether its response merely sounds plausible.

For grounding tasks, teams can reproduce production conditions and measure precision and recall for boxes or masks. Precision tracks how many identified results are correct; recall tracks how many relevant targets the system finds. Chart tests can include synthetic plots with difficult cases such as logarithmic scales or stacked series.

Safety testing must also consider misleading visual patterns and spurious correlations, where a model relies on an incidental feature instead of the relevant evidence. A system that chooses where to crop and zoom can develop confirmation bias if its plan repeatedly directs attention to the wrong region.

Organizations using a thinking mode will expect audit logs of its steps, or at least structured records of tool calls and bounding boxes. Broader independent evaluations will be needed to assess reliability and safety across these tasks.

An open option in a crowded AI race

Baidu's release arrives during a busy year for open models. DeepSeek introduced a large Mixture of Experts language model with hundreds of billions of total parameters and only a few dozen billion active per token. It reported strong general benchmark scores at a fraction of the training cost associated with closed systems.

Alibaba's Marco o1 explored reasoning through techniques including chain of thought fine tuning and planning guided by search. Those methods were aimed at open ended tasks that lack clear numerical rewards, where judging a good answer can be harder than checking a simple calculation.

Chinese laboratories, including Baidu and Alibaba, are increasingly releasing capable models under permissive licenses. Open access gives regional developers more choices and reduces dependence on proprietary services. It also places pressure on closed providers to justify premium prices through reliability, tools and integrations.

The performance gap between open models and closed systems has narrowed in many areas, especially specialized tasks such as chart analysis and document extraction. ERNIE adds another option in those categories, combining visual reasoning with an architecture intended to reduce the work required for each processing step.

Practical limits remain. An 80 GB GPU is beyond many consumer devices. The listed context window is 128K tokens, the amount of information the model can consider at once. That can help with long PDFs, but images, text recovered through optical character recognition and instructions all draw on finite capacity.

Deployment at scale also requires monitoring the balance between experts and delays caused by concentrated routing. Baidu's release brings open access, commercial licensing and a broad set of visual capabilities into one package; its usefulness for any particular business will depend on the accuracy, speed and reliability measured in that business's own workload.

Key Points

• Baidu released ERNIE 4.5 VL 28B A3B Thinking under Apache 2.0 for unrestricted commercial use.
• The model processes text, images, documents and video, activating about 3 billion of its 28 billion parameters per token.
• Baidu claims selected vision benchmark wins over GPT-5 High and Gemini 2.5 Pro, plus inference two to three times faster than comparable models with similar accuracy.
• Independent laboratories had not verified the benchmark charts as of November 12, 2025.
• Baidu says inference can run on one GPU with 80 GB of memory; the listed context window is 128K tokens.
• Thinking with Images supports zooming, cropping and repeated inspection, alongside grounding, image search and video understanding.
• Developer tools include ERNIEKit, Transformers, vLLM and FastDeploy, with training strategies named GSPO and IcePop.
• Production testing needs to cover task accuracy, operating conditions, routing performance, safety and records of the model's actions.