Last year, when chatting with friends about AI inference in the browser, nine out of ten would say "WebGPU can barely run small models." WebNN? What's that — a toy from some W3C community group.
This year, things have changed.
On January 22, 2026, W3C published an updated Candidate Recommendation Snapshot of the Web Neural Network API. What does the CR stage mean? The core design of the standard is frozen — no more major changes. The remaining work isn't "what to build" but "who can prove they can do it."
W3C requires two independent, interoperable browser implementations plus a publicly verifiable test suite before moving to Proposed Recommendation. In other words, WebNN has entered the stage where the exam proctor is asking candidates to hand in their papers.
---
From "toy" to "actually usable": what 2025–2026 added
I went through the Changelog, and the six items below are what convinced me WebNN is no longer a placeholder.
Transformer operator support. This is the biggest change. Previously WebNN's operator set basically stuck to pre-2019 CV model thinking — convolutions, pooling, fully connected layers; sufficient but not interesting. Now attention and quantizeLinear/dequantizeLinear have been added, meaning you can run quantized Transformer models in the browser. Not a toy demo — real inference producing meaningful results.
MLTensor API. WebNN and WebGPU can now share buffers without copying data back and forth. This matters more than it sounds — in the browser, the biggest cost of running models isn't compute, it's memory movement. Shuttle a 7B model's intermediate tensors between the JS heap and GPU buffers a few times and performance collapses. MLTensor opens up that path.
Accelerator selection. You can now explicitly specify the backend: GPU, NPU, or CPU. No more relying on the platform's "best guess." For debugging and performance tuning, this turns a black box into a configurable engine.
Worker support. Inference can now run in Shared Workers and Service Workers too. Previously WebNN (and WebGPU) contexts were bound to the main page — switch tabs and it's gone. Now you can run in a background thread, meaning inference in one tab while browsing normally in another.
Context loss handling. Recovery behavior after a GPU/NPU crash is now clearly defined. This isn't a performance feature — it's a mark of engineering maturity. An API only becomes a serious thing once it reaches the stage of "we need to think about what happens when things break."
Operator labels. Diagnostic labels for operators to ease debugging. A small feature, but it shows the API designers are starting to think about developer experience.
---
On the browser side: Chrome leads, the others are still warming up
In practice, WebNN's deployment looks like this:
Chrome / Edge: the most complete implementations, passing large numbers of WPT (Web Platform Tests). The Chromium team's investment in WebNN is serious — consistent with their WebGPU strategy: master the standard, pass the tests, then ship to users.
Backend coverage is decent too: DirectML on Windows, CoreML on macOS/iOS, NNAPI on Android, and OpenVINO on Intel platforms. Mainstream hardware is largely covered.
Safari / Firefox: slow to follow. Not news. Apple has always been conservative in its pace of web standards investment — WebGPU dragged on for ages too. The Firefox team has limited resources and can't prioritize this. Frankly, don't expect cross-browser consistency anytime soon.
What does this mean in practice? If your target users are Chrome users (in China, many people indeed only use Chrome or Chromium-based browsers), you can start experimenting now. If you're building a cross-browser product, you'll have to wait.
---
What you can do now
WebNN's practical capabilities have already surpassed the "proof of concept" stage:
- Running Stable Diffusion in the browser. Not the unusably slow kind — it genuinely produces decent images.
- Local Transformer inference. DistilBERT and small LLMs are fine. I've tried several demos; latency is acceptable.
- Real-time video analysis, speech-to-text. These always needed hardware acceleration, and WebNN offers a more specialized path than WebGPU.
- WebGPU as fallback. Automatically degrade to WebGPU when WebNN is unavailable — a low-risk progressive enhancement strategy for developers.
---
When will it be "truly usable"?
This question has two layers.
At the standards level, it's mature. The core design is frozen; no big changes coming. Vendor implementations and interoperability testing are the main storyline now.
At the ecosystem level, we still need to wait. The model conversion toolchain (ONNX → WebNN) is still maturing. To run a PyTorch model in the browser, you go through ONNX export → operator compatibility checks → WebNN format conversion, and every step can trip you up. Memory and performance bottlenecks for large models (7B+) in the browser are also unsolved — that's not an API problem, it's the browser sandbox's inherent memory limits.
Safari and Firefox's lag means that if you need a product with cross-browser consistency, now is not the time.
My judgment: from late 2026 to early 2027, WebNN will reach the "production ready" waterline. Provided Chrome keeps up its current pace of investment, and at least one more browser vendor (most likely Firefox; I don't hold much hope for Safari) follows.
---
A more interesting question
The most interesting part of this whole story, to me, isn't WebNN itself — it's that the addition of Transformer operators makes it genuinely useful for LLM inference for the first time.
This means the browser is transforming from a "rendering engine" into an "AI runtime." When we used to say "the browser is the new operating system," it was mostly a metaphor. Now the combination of WebGPU + WebAssembly + WebNN makes that metaphor feel less and less like a metaphor.
If this trend holds, a year or two from now, frontend engineers will care not just about DOM and CSS, but also quantization precision, KV Cache, and context windows. Will this transition happen? I don't know. But it's a direction worth watching.
---
*Posted on zhichai.net, June 2026, Shanghai*