BenchmarkXPRT Blog banner

Talking WebXPRT with Klaus Hinum of Notebookcheck

Justin Greene
By Justin Greene
on October 22, 2026

One of the keys to WebXPRT’s ongoing success as a browser performance benchmark has been a track record of consistent support from users of all types, including the tech press. From the early days of WebXPRT 2013 through today, analysts and device reviewers have used WebXPRT to evaluate all kinds of web-enabled devices.

Recently, we had the opportunity to talk WebXPRT with one of the most prolific tech reviewers in the industry: Klaus Hinum. Klaus is the co-founder and CTO of Notebookcheck, an international tech publication headquartered in Vienna, Austria that specializes in in-depth hardware reviews. Notebookcheck has been around since 2005 and has published deep-dive reviews on thousands of devices in over a dozen languages. That tremendous volume makes Notebookcheck one of the largest public sources of WebXPRT results.

Below, you’ll see the questions we asked Klaus in bold, followed by his answers.

Klaus, we appreciate you taking the time to answer our questions and share insights from your years of WebXPRT testing.

How does WebXPRT fit into your review process? What does a typical Notebookcheck test suite look like, and at what point in a review do you run WebXPRT?

WebXPRT is an integral part of all our extensive laptop, smartphone, tablet, and PC reviews. For laptops and PCs, we run the test using a semi-automated process. The benchmark is executed during the performance testing phase, though the exact sequence within that stage can vary slightly depending on the reviewer.

Roughly how many devices have you tested with WebXPRT, and across what categories? (For example, budget Chromebooks, gaming laptops, tablets, phones.)

Across all categories, ranging from smartwatches to Chromebooks, laptops, tablets, and phones. Everything that has a browser and runs WebXPRT. Here’s the breakdown of the number of devices for each version of WebXPRT:

  • WebXPRT 2013: 261
  • WebXPRT 2015: 556
  • WebXPRT 3: 2,929
  • WebXPRT 4: 1,682
  • WebXPRT 5: 115

How do you keep results consistent? What steps do you take with browser versions, power settings, thermals, or background apps before a run?

For PCs and laptops, we maintain standardized Chromium versions specifically for testing. Power profiles are configured to deliver optimal performance for each review, and the chosen settings are explicitly detailed in the article. For devices where the default browser cannot be standardized, such as smartphones or smartwatches, the browser variation is accounted for when evaluating the comparison results.

Have WebXPRT results or score trends ever surprised you? For example, a budget device outscoring a premium one, or a big gap between browsers on the same hardware.

In the past, Apple consistently led the field thanks to deep optimizations in Safari, and the introduction of Apple Silicon pushed those web performance scores even further. 

What does a browser benchmark tell your readers that other benchmarks don’t? Why does daily web performance matter for their buying decisions?

Since web browsing is one of the most common everyday tasks across almost all device categories, browser benchmarks are essential for assessing real-world performance that directly impacts user experience.

How useful are cross-platform comparisons to you? Do you compare scores across Windows, macOS, ChromeOS, Android, and iOS devices, and how do readers respond?

Yes, we regularly include cross-platform comparisons in our reviews. They provide vital context for our readers, especially in categories like tablets, where Android and Apple devices directly compete against one another on the same workload.

What makes a benchmark trustworthy enough to include in your reviews? (Repeatability, run time, transparency, ease of use, free access, etc.)

Key factors include ease of use, strong repeatability, and high transparency. The fact that WebXPRT is freely accessible allows our readers to easily run the benchmark on their own hardware and directly compare their results with our reviews.

WebXPRT 5 includes four AI-oriented workloads in the main test set, plus an optional AI image generation workload for high-end systems, all running locally in the browser. Does on-device AI performance matter to your readers yet? How do you expect that to change over the next few years?

On-device AI on laptops currently lacks a defining “killer feature,” whereas smartphones are increasingly relying on local AI processing for daily tasks. As developers integrate more local AI workloads into software, on-device benchmark performance will become significantly more important to our readers.

If you could add or change one thing in a future WebXPRT, what would it be?

A self-contained, easily deployable offline version would be a great addition. This would eliminate interrupted runs caused by network latency or internet connectivity issues.

Thanks again, Klaus! We’re taking note of your suggestion and will be looking into it.

If you’re involved with WebXPRT and, like Klaus, you’re willing to share your own WebXPRT-related thoughts here in the XPRT blog, please let us know.

Justin

Posted in AI, Apple, benchmark, Benchmark metrics, Benchmarking, BenchmarkXPRT, browser performance, Browser-based benchmarks, Chromebooks, Cross-platform benchmarks, Laptops, on-device AI, Performance benchmarking, Phones, results, Safari, Tablets, WebXPRT, WebXPRT 4 | Tagged , , , , , , , , , , , , , , , , , , , , , , , |

The WebXPRT 5 experimental AI image generation workload is here!

Justin Greene
By Justin Greene
on September 23, 2026

Last week, we announced that our first WebXPRT 5 experimental AI workload, Image Generation, was almost ready. Today, we’re excited to say we took it live on Monday! If you have a compatible system, you can try it right now.

If you missed our earlier posts about the experimental workload, here’s a quick recap of what it does, what you’ll need to run it, and how to get started. For more details about the workload, see the WebXPRT 5 Release Notes.

What does the workload do?

It uses on-device AI to generate images from text prompts. The Image Generation workload presents a text-to-image GenAI scenario. It uses Z-Image-Turbo, ONNX Runtime Web, Transformers.js, and WebGPU to produce four prompt-based images: one warm-up image and three successive scored images.

It reports an images-per-minute score. The workload score reflects how quickly your system generates the three scored images. Because GenAI models don’t produce identical output images, no two runs look exactly alike. Even so, the scores in our testing have been stable.

It does not affect the main WebXPRT score. The workload runs separately from WebXPRT 5’s seven core tests, and it does not affect your main WebXPRT 5 score or how the core tests run.

What do I need to run it?

This is a cutting-edge, resource-intensive workload, so it won’t run on many of the devices that WebXPRT usually supports, such as phones, tablets, and most Chromebooks. It’s best suited for late-model, high-performance laptops and desktops. Here’s what you’ll need:

  • A WebGPU-capable browser. Use a Chromium-based browser (Chrome 113+ or Edge 113+) or Safari 26+. On Qualcomm Snapdragon and other Windows-on-Arm systems, WebGPU is sometimes off by default, so you may need to manually enable it. If WebGPU is unavailable or disabled, the workload will not run.
  • A recent processor with at least 8 cores. Examples include Intel Core Ultra, AMD Ryzen (including Ryzen AI), Apple M-series, and Qualcomm Snapdragon X-class chips.
  • At least 32 GB of system memory. The model needs substantial memory to run smoothly.
  • Time and space for a large, one-time model download. The first time you run the workload on a system, it downloads the Z-Image-Turbo model (just under 6.5 GB) from the BenchmarkXPRT Hugging Face repository. Depending on your connection, that can take 5 to 20 minutes.

How do I run it?

Just a few clicks from the WebXPRT home page. Before you begin, set your device’s sleep timer to at least 30 minutes and close other apps and windows so background activity doesn’t affect your score.

Once the system is ready, go to www.WebXPRT.com, click Run WebXPRT 5, and click Continue. On the benchmark’s main page, click the Experimental AI workloads button below the main Start test button. On the Experimental AI workloads page, confirm the Image Generation checkbox is selected (by default, it should be) and click Start test. That’s it!

On the first run, WebXPRT 5 downloads the model and then automatically begins generating the images. After the download, each run takes about 5 minutes.

How do I use the results?

You’re free to use the experimental Image Generation workload’s images-per-minute score to compare the on-device, browser-based, text-to-image GenAI performance of devices. You’re also free to publish your scores, and we encourage you to do so!

As with all XPRT scores, it’s important to provide readers with adequate technical context by publishing relevant information about the test system’s hardware and software components. Part of this experimental workload process will include future workload version updates as AI models and technologies evolve, so it’s also important to note the workload’s version number when you publish scores (e.g., WebXPRT 5 Experimental AI Image Generation v5.0.1).

This is still an experiment, so please let us know how it goes!

The browser-based AI space is moving fast, and we expect to keep adjusting the workload as its underlying components evolve. We see this as an experiment we’re running together with the community, and we’ll all benefit from what we learn.

We want to hear how your testing goes, including workload failures on specific platforms or systems, potential bugs, performance comparisons, and ideas for improvement.

If you have any questions or comments about the Image Generation workload or our WebXPRT 5 experimental AI work, please let us know!

Justin

Posted in AI, AMD, Apple, Arm, benchmark, Benchmarking, BenchmarkXPRT, BenchmarkXPRT development community, browser performance, Browser-based benchmarks, Chrome, Chromium, Cross-platform benchmarks, Future of performance evaluation, GenAI, Hugging Face, image generation, Intel, JavaScript, Microsoft Edge, on-device AI, ONNX Runtime Web, ORT Web, Performance benchmarking, Qualcomm, results, Ryzen AI, Safari, Snapdragon X Elite, text-to-image, Transformers.js, Web AI, WebGPU, WebXPRT, WebXPRT 5, Z-Image-Turbo | Tagged , , , , , , , , , , , , , , , , , , , , , , , , , , |

The WebXPRT 5 experimental AI image generation workload is almost here!

Justin Greene
By Justin Greene
on September 16, 2026

In last week’s blog post, we discussed how our work on experimental AI workload candidates for WebXPRT 5 led us to focus on a single initial use case: image generation. Today, we’re excited to announce our plans to publish the experimental image generation workload next week!

If you didn’t get a chance to read last week’s post about what the workload will do and what to expect when you run it on your system, you’ll find a recap of some of that information below, along with new details and answers to common questions.

What does the workload do?

Executes prompt-based image generation. The Image Generation workload presents a text-to-image GenAI scenario. It uses Z-Image-Turbo, ONNX Runtime Web, Transformers.js, and WebGPU to produce four prompt-based images: an initial warm-up image and three successive scored images.

Produces an images-per-minute score. The test reports a single score, measured in images per minute. Due to the nature of GenAI model technology, the workload unavoidably generates slightly different image outputs each time, so no two runs look exactly alike. That said, the outputs are stable enough to produce consistent scores with low variance.

How will I run the workload?

Two clicks from the main WebXPRT 5 Start test page. To run the experimental Image Generation workload, you’ll click the Experimental AI workloads button below WebXPRT 5’s normal Start test button (see Figure 1 below). On the Experimental AI workloads page, make sure that the checkbox for the Image Generation workload is checked, and click Start test (see Figure 2 below).

Will I need to download anything?

Yes. After you click Start test, the workload will initiate a download of the Z-Image-Turbo model from the BenchmarkXPRT Hugging Face repository. The model size is just under 6.5 GB, so the download could take 5 to 15 minutes depending on your internet speed. You’ll need to download it only the first time you run this test on a system.

WebXPRT 5 will run the workload’s tasks automatically once the download is complete.

Figure 1: The WebXPRT 5 Start test page, showing the new Experimental AI workloads button.
Figure 2: The WebXPRT 5 Experimental AI workloads page, showing the Image Generation workload.

What else do I need to know about the workload?

Typical WebXPRT 5 runs involve a few clicks and a few minutes on almost any system. Running this new experimental AI workload will be a different, but still straightforward, experience. Here are a few points to keep in mind:

  • The workload will be completely optional and will run separately from WebXPRT 5’s main tests. You do not have to run the Image Generation workload. We just think you’ll want to. No matter what, the main WebXPRT 5 testing experience and scores will not change.
  • It won’t run on everything. WebXPRT has a well-earned reputation for running on just about anything with a browser. Because the Image Generation workload is a cutting-edge, resource-intensive, browser-based, experimental performance test that has interfaces down to system hardware, it won’t run on that same broad range of devices.
  • We recommend testing on systems with at least 32 GB of memory. This workload is best suited to late-model, high-performance laptops and desktops. We’ll share more information about minimum hardware specs on launch day.
  • It takes longer to run. This is a demanding workload, so you should expect longer runtimes than those of a typical WebXPRT 5 test. After the initial model download, each run may take about 5 to 10 minutes.

This is very much an experiment

The AI world in general, and the browser-based AI world in particular, are changing at astonishing speeds. We expect to have to make future adjustments to the workload as the underlying technical components evolve, but we wanted to give you a way to interact with emerging AI technologies in a benchmark environment. We view this as an experiment that we’re running together, and we’ll all benefit from what we learn.

We want your feedback

After the launch, we want to hear your thoughts on the Image Generation workload, along with any observations you have from testing. We’re interested in hearing about workload failures on specific platforms and systems, potential bugs, performance comparisons, and suggestions for improvement. We welcome it all!

For now, if you have any questions or comments about the upcoming workload or our WebXPRT 5 experimental AI work, please let us know!

Justin

Posted in AI, benchmark, Benchmark metrics, BenchmarkXPRT, BenchmarkXPRT development community, browser performance, Browser-based benchmarks, Cross-platform benchmarks, GenAI, Hugging Face, image generation, JavaScript, Laptops, on-device AI, ONNX Runtime Web, ORT Web, Performance benchmarking, text-to-image, Transformers.js, Web AI, Web-based testing, WebGPU, WebXPRT, WebXPRT 5, Z-Image-Turbo | Tagged , , , , , , , , , , , , , , , |

Our first WebXPRT 5 experimental AI workload is taking shape!

Justin Greene
By Justin Greene
on September 10, 2026

We know that many of you have been patiently waiting for more details on an experimental, browser-based local AI workload for WebXPRT 5. After weighing several candidates and evaluating their current suitability for a browser benchmark, we’ve narrowed our focus to a single initial use case: Image generation! We’re not yet ready to announce a launch date for the new workload, but our development and testing work has advanced far enough that we can now share some details.

Generating images with AI is one of the most popular local AI activities today, which makes this use case a good fit for the kind of demanding, relevant, and forward-looking tests we want to offer in the experimental workload section. In a previous blog post, we discussed some of the decision points we faced as we explored our options. In the sections below, we provide more information about what the Image Generation workload is likely to do, explain how we answered those exploratory questions for this first experimental AI workload, and let you know what to expect when you run the test.

About the Image Generation workload

Prompt-based image generation. The text-to-image GenAI workload produces an initial warm-up image and then three successive scored images. The test produces a single score using a metric of images per minute. Due to the nature of GenAI model technology, the workload unavoidably generates slightly different image outputs each time, so no two runs look exactly alike. That said, the outputs are consistent enough to produce consistent scores with low variance.

The underlying Web AI frameworks. The workload draws on two web AI frameworks:

  • ONNX Runtime Web (ORT Web) is the JavaScript (JS) and web deployment version of ONNX Runtime, Microsoft’s cross-platform inference engine for models in the Open Neural Network Exchange (ONNX) format. With available tools, users can convert models from several AI frameworks to ONNX, which ORT Web can then use. ORT Web allows developers to leverage the compatibility of ONNX-formatted ML models and reduce the number of expensive CPU-to-GPU data transfers that multi-stage image generation pipelines would otherwise require.
  • Transformers.js is a JS library for running pretrained transformer AI models in browsers and JS applications. The library downloads compatible models from the Hugging Face Hub, caches them locally, and runs them on the CPU through WebAssembly or on the GPU through WebGPU.

The web API. WebGPU enables web-based applications, such as image-based GenAI and inference workloads, to directly access the graphics rendering and computational capabilities of a system’s GPU. WebGPU builds on modern native graphics APIs such as Direct3D 12, Metal, and Vulkan, and offers a mature and broadly available path for running demanding models in current browsers. WebNN is a higher-level API designed to tap available AI hardware, including GPUs and NPUs, but it is still maturing and carries limited browser and model-operator support. For now, WebGPU is the best path for WebXPRT 5.

The workload’s image generation model. To power the workload, we selected Z-Image-Turbo, an Apache-2.0-licensed, 6-billion-parameter text-to-image model in the Z-Image family. Its scalable single-stream diffusion transformer combines text, visual semantic, and VAE tokens in one sequence for parameter efficiency. That relatively streamlined process, along with a derived ONNX model that runs through WebGPU, makes Z-Image-Turbo a practical choice for interactive use in a browser.

What to expect

Typical WebXPRT 5 runs involve a few clicks and a few minutes on almost any system. Running this new experimental AI workload will be a different, but still straightforward, experience. Here are a few points to keep in mind:

  • The workload will be completely optional and will run separately from WebXPRT 5’s main tests. Your main WebXPRT 5 scores and testing experience will not change.
  • It won’t run on everything. WebXPRT has a well-earned reputation for running on just about anything with a browser. Because the Image Generation workload is a cutting-edge, resource-intensive, browser-based, experimental performance test, it won’t cover that same broad range of devices. It will best suit late-model, high-performance laptops and desktops. We’ll share more about minimum hardware specs before the launch.
  • It will require a large one-time model download. You will need to download the Z-Image-Turbo model from the BenchmarkXPRT Hugging Face repository. The total size of the model download is just under 6 GB, but you need to download it only the first time you run this test.
  • It takes longer to run. Expect longer runtimes than a typical WebXPRT 5 test. Depending on your internet speed, the initial model download can take 5 to 20 minutes. After that, each run may take about five minutes.

We’ll update you here in the blog with more information on all of these fronts as development continues.

We’ll keep you posted

We’re glad to be able to provide more concrete information about our experimental AI workload progress, and we think the wait will prove worthwhile. If we don’t hit unexpected technical snags, we hope to be able to soon share a launch date.

As always, we welcome your input. If you have thoughts on the Image Generation workload, the frameworks and APIs behind it, or anything else related to WebXPRT 5 or our experimental AI work, please let us know!

Justin

Posted in AI, Apache 2.0, benchmark, Benchmarking, BenchmarkXPRT, BenchmarkXPRT development community, browser performance, Browser-based benchmarks, Cross-platform benchmarks, Direct3D 12, Future of performance evaluation, GenAI, Hugging Face, image generation, JavaScript, Machine learning, Metal, Microsoft, ONNX Runtime Web, ORT Web, Performance benchmarking, text-to-image, Transformers.js, Vulkan, Wasm, Web AI, web API, Web-based testing, WebAssembly, WebGPU, WebNN, WebXPRT, WebXPRT 5, Z-Image-Turbo | Tagged , , , , , , , , , , , , , , , , , |

Four key reasons you should move to WebXPRT 5

Justin Greene
By Justin Greene
on August 6, 2026

We’re excited to see continued growth in the number of tech press outlets, OEM labs, and independent testers that have added WebXPRT 5 to their suite of go-to benchmark tools for hardware analysis. We know that for tech enthusiasts who invest significant time in carefully evaluating new gear, the decision to use WebXPRT 5 reflects a high degree of confidence in the benchmark’s relevance in a rapidly changing tech marketplace.

For many WebXPRT users, WebXPRT 5 adoption is also the natural evolution of years of reliable WebXPRT 4 testing. WebXPRT 4 has been phenomenally successful, with over 725,000 completed test runs to date. While we love WebXPRT 4 and are grateful to have earned the trust of so many users, WebXPRT 5 is the future of cross-platform browser performance benchmarking in the AI age. At a glance, it may seem like WebXPRT 5 is simply a version of WebXPRT 4 with an updated UI and rescaled scoring. Dig deeper, though, and you’ll see that WebXPRT 5 represents a significant upgrade from WebXPRT 4—one that is well-positioned for the future.

If you haven’t yet shifted your testing from WebXPRT 4 to WebXPRT 5, there are several key reasons why we think it makes a lot of sense to go ahead and make the remarkably easy and pain-free transition.

WebXPRT 5 offers more AI workloads, with full transparency

We added more timed AI tasks with WebXPRT 5, along with the kind of technical transparency that’s vital for in-depth performance analysis or defending test methodologies in published collateral. Four of WebXPRT 5’s seven workloads utilize AI capabilities, and we provide detailed information about the workloads’ actual models and runtimes: Video Background Blur uses an ONNX human-segmentation model with OpenCV.js (WASM); Detect Faces uses a ResNet-based Caffe model; Image Classification uses SqueezeNet v1.0 on ImageNet images; and Document Scan uses tesseract.js v6.0.1 OCR with LSTM. WebXPRT’s ability to measure how well devices handle those types of workloads has significant value and relevance for modern performance analysis.

We also offer free access to the WebXPRT 5 source code, so you don’t have to take our word for how everything works. You can look under the hood to your heart’s content.

WebXPRT 5 provides updated, higher-resolution content and longer operations

With WebXPRT 5, we updated much of the benchmark’s workload content to reflect changes in everyday technology, such as upgrading most photos in the photo-processing workloads to higher resolutions. One example is the Photo Effects workload, which now applies an Emboss filter to photos with resolutions up to 4928 x 3264. Higher-resolution inputs both reflect today’s better phone (and other) digital cameras and can stress modern devices more meaningfully, reducing score compression at the top of the range.

In addition, many of WebXPRT 4’s existing workloads no longer challenge cutting-edge consumer hardware as much as many of us would like. Testing labs recognized this and asked us for longer and more demanding workloads. We developed workloads for WebXPRT 5 that are accessible enough that you can run them on a broad range of devices yet challenging enough to allow performance differentiation among high-end systems.

You can directly compare your results to those from major tech press outlets

Several global tech press outlets have already transitioned to testing with WebXPRT 5 or are in the process of doing so, including PCMag, Notebookcheck, ComputerBase (Germany), XFastest (Taiwan), eTeknix, BenchLife.info, and ITC.ua (Ukraine). As the hardware review industry shifts to WebXPRT 5 for browser performance evaluation in the age of AI, it makes sense to test the newest gear with the newest tests and have the ability to quickly compare your results to thousands of others.

Even more WebXPRT 5 AI is on the way!

In past blog posts, we’ve discussed our goal of developing one or more experimental WebXPRT workloads. Experimental workloads are optional, and their scores stay separate from the main overall score, so we can add or adjust the tests without affecting comparability. The experimental workload approach gives us the flexibility to put cutting-edge measurement tools in your hands—even if those tools won’t yet run on every platform that supports the main WebXPRT tests. This setup is especially well-suited for the whirlwind of change and technological advancement that we’re seeing with AI.

For instance, there’s a real need for tests that target local, browser-based AI workloads like LLM inference. As the foundational technologies of Web AI mature—such as WebGPU and eventually WebNN—we can work on developing those types of workloads and adding them as experiments without inconveniencing existing WebXPRT 5 users.

Though we are actively working on experimental AI workloads, we’re not yet ready to share details here in the blog—but we are very excited about what’s to come! If you make the transition to testing with WebXPRT 5, you’re going to have access to pioneering experimental workloads right out of the gate—workloads that won’t be available with any other browser performance benchmark.

If you’ve been interested in WebXPRT 5 but you haven’t yet made the switch, we hope this blog post will encourage you to give it a try.

If you have any questions or comments about the differences between WebXPRT 4 and WebXPRT 5, please let us know!

Justin

Posted in AI, benchmark, Benchmark metrics, Benchmarking, BenchmarkXPRT, BenchmarkXPRT development community, browser performance, Browser-based benchmarks, Cross-platform benchmarks, Future of performance evaluation, image classification, ImageNet, inference, OCR, on-device AI, ONNX Runtime Web, Performance benchmarking, ResNet-50, SqueezeNet, Transformers.js, Wasm, Web AI, web API, WebNN, WebXPRT 4, WebXPRT 5 | Tagged , , , , , , , , , , , , , , , , , , , , , |

WebXPRT 5: AI tests now, lots of room for growth

Justin Greene
By Justin Greene
on July 9, 2026

In past blog posts, we’ve discussed our goal of developing one or more experimental WebXPRT workloads focused on local, browser-side AI technologies. While many of us regularly interact with cloud-based AI apps and services through a browser, on-device AI capabilities are growing rapidly, and we want WebXPRT to continue to evolve with them.

There are several driving factors behind that growth. Web API technologies keep maturing, giving browsers direct access to the hardware they need for real inference work. Advanced GPU and NPU technology is now widely available in consumer devices, so the local computing power necessary to run AI applications on-device is in reach for many users. And for many organizations, there are compelling reasons to execute increasingly vital work like LLM inferencing and agentic coding tasks on local machines—such as data privacy, regulatory compliance, and cost control.

The reasons for the experimental workload approach

The expansion of on-device AI is exactly the type of shift we built the experimental workload concept to capture. As we shared when we first announced the WebXPRT 5 workload lineup, an experimental workload section gives us the flexibility to put cutting-edge measurement tools in users’ hands—even if those tools won’t yet run on every platform WebXPRT has traditionally supported. Experimental scores stay separate from the main overall score and are completely optional, so we can add tests without affecting comparability or asking anyone to retest. That approach maintains WebXPRT’s strengths while preparing the benchmark for the future—and giving all of us valuable information today.

The AI functions that WebXPRT 5 measures today

WebXPRT 5 already includes four workloads that utilize AI capabilities: Video Background Blur with AI, Detect Faces with AI, Image Classification with AI, and Document Scan with AI. These workloads use machine learning—computer vision and OCR models such as a Caffe-based face detector, SqueezeNet for image labeling, and an LSTM-based OCR engine. WebXPRT’s ability to measure how well devices handle those types of workloads has real value, and it reflects the kinds of light browser-side inference tasks that have been in widespread use for a while.

We recognize, though, that there’s a clear need for more demanding local, browser-based AI workloads—especially LLM inference. We’re targeting that need with our experimental work. Like pretty much everyone else, we’re also developing in the midst of an incredibly dynamic technical environment. We want to purposefully move forward without sacrificing WebXPRT’s stability and reliability for the sake of expedience.

The main decisions we face

Choosing a Web AI framework. We’re still researching our open-source framework options, including candidates like ONNX Runtime Web, Transformers.js, MediaPipe, and TensorFlow.js. The ground here continues to shift. For example, Transformers.js v4 now supports a WebGPU backend and spans a very broad range of model architectures. So, one of our ongoing challenges is picking a durable foundation.

Choosing a web API. Of the primary options we’re investigating, WebGPU now has the broadest browser support (Chrome, Edge, and partial support in Firefox and Safari). WebNN remains the most promising option in the long term because it can directly target NPUs, but it’s still not ready for production—its W3C spec only reached Candidate Recommendation status in early 2026, and browser support outside of flagged, experimental builds isn’t there yet. Our web API outlook hasn’t changed much from before: WebGPU is the most practical path today, and WebNN may be an exciting possibility for tomorrow.

Choosing and sizing workloads. We’ll ideally find workloads demanding enough to genuinely stress new hardware, but light enough to run on slightly older gear without forcing huge model downloads or overextending the test’s runtime. The sweet spot for browser inference today tends to be small, quantized models, and memory ceilings and cold-start downloads are real constraints. Striking the right balance is another part of the challenge we’re working through.

We appreciate your patience

We’ve been talking about experimental WebXPRT AI workloads for a while. While we wish we already had everything worked out, we think the end product will be worth the wait. We appreciate your patience as we work through the details, and we’ll keep updating you here in the blog as we make progress.

As always, we’re open to suggestions. If you have ideas for a browser-based AI workload scenario, a framework or API you think we should weigh, a browser-based AI application you want us to consider, or any other related thoughts, please let us know!

Justin

Posted in AI, benchmark, Benchmarking, BenchmarkXPRT, browser performance, Browser-based benchmarks, Caffe, Chrome, Chromium, Collaborative benchmark development, computer vision, Cross-platform benchmarks, Future of performance evaluation, GPU, HTML5, image classification, inference, JavaScript, large language models, MediaPipe, Microsoft Edge, NPU, OCR, ONNX Runtime Web, Performance benchmarking, Performance of computing devices, Safari, SqueezeNet, TensorFlow, Wasm, Web AI, web API, Web-based testing, WebAssembly, WebGPU, WebNN, WebXPRT, WebXPRT 5 | Tagged , , , , , , , , , , , , , , , , , , , , , , , , , |

Check out the other XPRTs:

Forgot your password?
© Principled Technologies, Inc. 2026   |   Privacy policy

Cookie Notice: Our website uses cookies to deliver a smooth experience by storing logins and saving user information. By continuing to use our site, you agree with our usage of cookies as our privacy policy outlines.

x