A 12B decision classifier in the browser. A state, a question and its options in; one probability per option out, from a single forward pass on your GPU.
Jev-Omni is a Gemma 4 12B model with a 256-way decision
head. @ai-ecoverse/jev-omni.js runs it on WebGPU through onnxruntime-web, for text and images.
Try it
Read this before you press the button. The first load downloads 13.6 GB
from Hugging Face and keeps it in this browser's
cache, so it downloads once. The model needs about 16 GB of GPU memory, and the tab holds the files in memory
while they are uploaded to the GPU: plan on a machine with 64 GB or more of RAM or unified memory (it was tested
on an Apple M4 Max with 128 GB). It needs WebGPU: a recent Chrome or Edge.
Checking for WebGPU…
One probability per option appears here, with the time it took.
Prediction JSON
Status
Text and images work in Chrome on WebGPU and match the full-precision model (below). The browser bundle is
13.6 GB and needs about 16 GB of GPU memory, so it is for machines like a large-memory Apple Silicon Mac. The
weights are on Hugging Face. Video and audio are
still to come.
Results
Chrome on WebGPU, Apple M4 Max, int8 weights with fp32 activations, against the fp32 PyTorch model.
Set
Accuracy, browser / reference
Answers changed
Mean / max |Δp|
DecisionBench medium, 293 text questions
86.35% / 86.35%
0
0.0019 / 0.131
kev.js vision-v1, 106 image questions
98.1% / 99.1%
1
0.0014 / –
kev.js vision-v2, 128 harder image questions
85.2% / 85.2%
1
0.0037 / 0.065
Latency on a quiet GPU: about 1.2–1.6 s per image question (31 ms for the image step), about 7 s for a text
question under 2k tokens, longer for long prompts (up to about 8k tokens with onnxruntime-web 1.30). Details: text,
images,
feasibility.
API
import * as ort from "onnxruntime-web/webgpu";
import { loadJevOmni } from "@ai-ecoverse/jev-omni.js";
const jev = await loadJevOmni("https://huggingface.co/ai-ecoverse/jev-omni.js/resolve/main/jev-omni", { ort });
const res = await jev.predict({
state: "The meeting starts at 10 AM. It is now 9 AM.",
question: "Has the meeting started?",
options: ["Yes", "No"],
// image: { width, height, data } // optional RGBA pixels
});
res.prediction; // the likeliest option
res.probabilities; // { Yes: …, No: … }
2 to 256 options per question; quality is established up to 20. The loader checks every file's SHA-256 against
the manifest and keeps the files in Cache Storage under the manifest's revision, so later loads read from disk
and a new revision replaces the old one. Run it in a Web Worker, as this page does. With onnxruntime-web 1.30,
attention needs an n×n buffer per head, which caps prompts at about 8,000 tokens on a GPU with 4 GB buffers (as
on Apple Silicon) and less on GPUs with smaller ones; predict rejects longer prompts with a clear
error. Long prompts are also slower than with the sliding-window attention that later onnxruntime-web versions
bring.
How it works
Decoder: Jev-Omni's own bf16 weights, exported with a vendored draft of the onnxruntime-genai model builder that supports Gemma 4, as a hidden-state graph with int8 weights and fp32 activations. The 8 global-attention layers run in fp16 to fit Apple GPUs' workgroup memory.
Head: the last position's hidden state, normalized, through one linear layer to 256 logits, masked to the number of options.
Images: Gemma 4 has no separate vision tower. Images become 48 px patches through a small embedder (200 MB, fp32), ported to JS with pixel checks against Hugging Face's processor, and tokens of one image attend to each other in both directions.
Compared with kev.js and cua-s1.js
In decision-vision-bench, Jev-Omni ties
Kev-4B vision on image questions, at 2.5× the download. It is clearly the strongest of the three on long text
decisions (92.5% on the DecisionBench slice). For images in the browser, use
kev.js; for long text decisions on a big machine, use this.