hjLabs AI Playground
Starting the on-device runtime…
Qwen3 4B is a 4B-parameter ONNX model that generates text a token at a time in response to a prompt, streaming its reply as it decodes. On this page it runs entirely inside your browser: the 2.9 GB download comes straight from the Hugging Face repository “onnx-community/Qwen3-4B-ONNX” into your browser cache, and every inference after that executes on your own GPU through WebGPU. This is one of the largest models the playground offers, so treat the first load as a deliberate decision rather than something to try on mobile data. It wants a discrete GPU with headroom, and it is worth requesting persistent storage before you start so the browser does not evict it later. The model is multilingual, so it is not restricted to English input.
The largest model this site can run: 4 billion parameters in a browser tab. Needs a GPU with shader-f16 and a few GB of VRAM. In practice people reach for a model like this for drafting, rewriting, summarising a pasted passage, or simply having a conversation that no server ever sees.
Available precisions: q4f16. The runtime picks the first entry your device can actually execute and falls back down the list. The f16 build only loads on adapters that report the WebGPU "shader-f16" feature, which many desktop GPUs — including NVIDIA cards under Linux Chrome — do not, so a non-f16 fallback is always listed after it. Hugging Face records about 877 downloads a month. It is a specialist choice rather than a default, which is exactly why it is worth trying when the popular model gets your particular input wrong.
Because there is no inference server behind this page, there is nothing to rate-limit and nothing to bill: the only cost of running Qwen3 4B here is the one-off 2.9 GB download, and once that is cached the page keeps working with the network switched off.
qwen3 4bqwen3 4b onnxqwen3 4b browserrun qwen3 4b locallyonnx-community/qwen3-4b-onnxchat modelqwen3 4b webgpuqwen3 4b download sizefree chat modeltransformers.js model
About 2.9 GB for the q4f16 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. onnx-community/Qwen3-4B-ONNX 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.