hjLabs AI Playground
Starting the on-device runtime…
DeBERTa v3 XSmall NLI is a 22M-parameter ONNX model that scores text against categories you invent at query time by reframing the problem as natural-language inference. On this page it runs entirely inside your browser: the 98 MB download comes straight from the Hugging Face repository “Xenova/nli-deberta-v3-xsmall” into your browser cache, and every inference after that executes on your own GPU through WebGPU. That is a one-off download of well under a minute on a normal connection, and small enough that integrated graphics handle it without complaint. It is an English-language model; expect quality to fall off sharply on other languages.
Best accuracy per megabyte for custom labels. Remember each label costs a separate forward pass. In practice people reach for a model like this for intent routing, topic labelling and triage where there is no labelled training set and never will be.
Available precisions: q8 (8-bit) and fp32 (full precision). q8 comes first because it roughly halves the download at a quality cost you will struggle to measure on this class of model; fp32 is there when you want the reference numbers. It is this playground's default pick for the task, and Hugging Face records about 7k downloads a month for the repository — a good sign that it is maintained and that the ONNX export is not abandoned.
Because there is no inference server behind this page, there is nothing to rate-limit and nothing to bill: the only cost of running DeBERTa v3 XSmall NLI here is the one-off 98 MB download, and once that is cached the page keeps working with the network switched off.
deberta v3 xsmall nlideberta v3 xsmall nli onnxdeberta v3 xsmall nli browserrun deberta v3 xsmall nli locallyxenova/nli-deberta-v3-xsmallzero-shot text labels modeldeberta v3 xsmall nli webgpudeberta v3 xsmall nli download sizefree zero-shot text labels modeltransformers.js model
About 98 MB for the q8 build, including the tokenizer and config files the runtime also has to fetch — this is the whole payload, not just the weights file. It is downloaded once and then served from the browser Cache Storage API.
Yes, through the WebAssembly fallback, which is slower but works anywhere. WebGPU (Chrome or Edge 113+, Safari 26+) is what makes it feel interactive. On WASM, prefer the smaller quantized build.
No. Xenova/nli-deberta-v3-xsmall is a public repository, so your browser fetches the ONNX files directly over HTTPS. There is no account, no key and no server in between.
No. The model weights are downloaded from Hugging Face to your browser once, and every inference after that runs on your own GPU through WebGPU. Your text, images and audio are never sent anywhere — hjLabs.in has no inference server and no way to see your input.