A model is still running

A model is still running or downloading. Leaving will stop the current operation. Stay on this page to let it finish.

Skip to content
LocalAI.toolsBeta

Join the waitlist

Get an email when the new apps and model runs launch.

No account needed.

Open Jev DeBERTa V3 Large — Typed Decisions

Answer structured questions about English text by scoring categories, ordered choices, or yes/no options. Returns calibrated option distributions without generating prose; an independent open model, not TypeSafe AI's Jev.

Browser inferencePowered by WebAI · Private · Unlimited runs
About this model

Answer structured questions about English text by scoring categories, ordered choices, or yes/no options. Returns calibrated option distributions without generating prose; an independent open model, not TypeSafe AI's Jev.

Created by
Kotoba Labs
License
Apache-2.0
Model released
Unknown
Source repository created
Sep 18, 2026
Model weights · Q4F16
348.4 MB

Best for

  • Route or classify English inputs among an application-provided set of choices.
  • Estimate ordered ratings and yes/no probabilities for human-reviewed workflows.
  • Prototype deterministic structured decision interfaces in the browser.

Keep in mind

  • This independent reproduction is not TypeSafe AI's proprietary Jev model and is not affiliated with TypeSafe AI.
  • The source reports 0.854 in-domain accuracy but 0.690 on new instructions and option sets; application-specific evaluation and recalibration are required.
  • It is English-only, has a 512-token total context, and was trained on banking support, movie reviews, and Wikipedia yes/no data.
  • Unseen ordered scales are a weak case, and out-of-domain confidence may be over-confident.
  • It cannot generate explanations, and its probabilities do not establish truth, safety, or suitability for consequential decisions.
  • The model is large for browser use: even the smallest artifact is roughly 348 MB before shared tokenizer files.
  • The isolated public-API audit did not retain the q4f16 artifact in Chrome's Cache API, so some browser profiles may redownload it between sessions.

Processed on your deviceInference runs in this browser. Input/output sharing requires consent for each run. Basic run logs include device information and approximate connection location. Model and runtime files download from external hosts. Saved downloads may be cleared by your browser.

Sources & technical details
Original model
com-kotobalabs/open-jev-deberta-v3-large
Release verification
Unknown
Release evidence
Exact original checkpoint public release date has not been verified. Repository creation, training, paper, artifact conversion and update dates are not release evidence.
Source updated
Sep 18, 2026
Browser model files
onnx-community/open-jev-deberta-v3-large-ONNX
Files updated
Sep 20, 2026
Artifact revision
7c79f25b5ac496089f448a969c801872ad59d31c

Input

TEXT

Sample result

Open Jev DeBERTa V3 Large — Typed Decisions
Advanced settings 3
Maximum total encoded tokens; requests whose questions do not fit are rejected. Lower values may process less text but truncate relevant context sooner. Higher values retain more context, within the model limit, and can need more memory and time.
Maximum state tokens retained before questions are appended. Lower values discard more state context. Higher values preserve more context within the model limit, potentially using more processing time and memory.
Softmax calibration temperature; 1.05 is fitted by the source model. Lower values favor likely tokens and more repeatable wording. Higher values increase variation and can increase mistakes. It affects sampling only; some runtimes use zero for greedy decoding.

Output

Your result will appear here

Add input or try the sample, then run the model.

Benchmark evidence

Saved worker evidence is separate from benchmarks run on your device. It does not guarantee speed or quality on every browser.

Open Jev DeBERTa V3 Large — Typed Decisions

8 configurations tested
Passed6
Failed2
TestedSep 22, 2026
RuntimeTransformers.js 4.2.0

Quality criteria

All checked-in typed questions must return finite normalized distributions, the expected top choice/yes-no decision, and an ordered-score value within the fixture tolerance; outputs must be non-degenerate.

PrecisionDeviceResult
Q4F16GPU · WebGPUAcceleratedPassedFast
FP32GPU · WebGPUAcceleratedPassedFast
Q4GPU · WebGPUAcceleratedPassedModerate
FP16GPU · WebGPUAcceleratedPassedModerate
FP32CPU · WASMPassedModerate
FP16CPU · WASMPassedSlow
Q4CPU · WASMFailed—
Q4F16CPU · WASMFailed—
Test setup & sample details
Quality gate
All checked-in typed questions must return finite normalized distributions, the expected top choice/yes-no decision, and an ordered-score value within the fixture tolerance; outputs must be non-degenerate.
Tested
2026-09-22T14:47:49.157Z
Environment
Isolated headless Google Chrome 151 on macOS; Transformers.js 4.2.0
Runtime
Transformers.js 4.2.0
Artifact revision
7c79f25b5ac496089f448a969c801872ad59d31c
Fixture
{
  "file": "open-jev-deberta-v3-large.json",
  "url": "https://f005.backblazeb2.com/file/hubters-webai-public/assets/benchmarks/open-jev-deberta-v3-large.json",
  "size": 819,
  "sha1": "bf1e5a03646050ac2c9b0fa51bf8f01b72f95ece",
  "source": "Adapted from the pinned source model card usage example.",
  "sourceUrl": "https://huggingface.co/com-kotobalabs/open-jev-deberta-v3-large/blob/19bf9a64815add579fbf6c907bef584d9277a8e4/README.md#use",
  "license": "Apache-2.0",
  "licenseUrl": "https://www.apache.org/licenses/LICENSE-2.0.txt",
  "preprocessing": "Pinned SentencePiece tokenizer; state truncates to 256 tokens; questions and options use the artifact's marker-token span interface."
}
Rerun when
artifact revision change; runtime version change; fixture or quality-gate change; browser/backend behavior change
Per-case quality & failure details
[
  {
    "precision": "fp32",
    "device": "wasm",
    "status": "pass",
    "quality": {
      "normalized": true,
      "finite": true,
      "expected": true,
      "nondegenerate": true,
      "grade": "typed-decision-pass"
    },
    "runtime": {
      "loadMs": 182152.3,
      "inferenceMs": 1475.1,
      "totalMs": 185097.3,
      "grade": "slow",
      "questionsPerSecond": 2.0338
    }
  },
  {
    "precision": "fp32",
    "device": "webgpu",
    "status": "pass",
    "quality": {
      "normalized": true,
      "finite": true,
      "expected": true,
      "nondegenerate": true,
      "grade": "typed-decision-pass"
    },
    "runtime": {
      "loadMs": 181271.3,
      "inferenceMs": 445.4,
      "totalMs": 183191.6,
      "grade": "good",
      "questionsPerSecond": 6.7355
    }
  },
  {
    "precision": "fp16",
    "device": "wasm",
    "status": "pass",
    "quality": {
      "normalized": true,
      "finite": true,
      "expected": true,
      "nondegenerate": true,
      "grade": "typed-decision-pass"
    },
    "runtime": {
      "loadMs": 110955.2,
      "inferenceMs": 2716.9,
      "totalMs": 115036.4,
      "grade": "slow",
      "questionsPerSecond": 1.1042
    }
  },
  {
    "precision": "fp16",
    "device": "webgpu",
    "status": "pass",
    "quality": {
      "normalized": true,
      "finite": true,
      "expected": true,
      "nondegenerate": true,
      "grade": "typed-decision-pass"
    },
    "runtime": {
      "loadMs": 97316.4,
      "inferenceMs": 965.6,
      "totalMs": 99586.3,
      "grade": "good",
      "questionsPerSecond": 3.1069
    }
  },
  {
    "precision": "q4",
    "device": "wasm",
    "status": "fail",
    "quality": "Not embedded",
    "runtime": {
      "loadMs": null,
      "inferenceMs": null,
      "totalMs": 64089.3,
      "grade": "failed"
    },
    "failure": "Could not find a WASM implementation for GatherBlockQuantized(1)."
  },
  {
    "precision": "q4",
    "device": "webgpu",
    "status": "pass",
    "quality": {
      "normalized": true,
      "finite": true,
      "expected": true,
      "nondegenerate": true,
      "grade": "typed-decision-pass"
    },
    "runtime": {
      "loadMs": 58865.6,
      "inferenceMs": 745.3,
      "totalMs": 61101.1,
      "grade": "good",
      "questionsPerSecond": 4.0252
    }
  },
  {
    "precision": "q4f16",
    "device": "wasm",
    "status": "fail",
    "quality": "Not embedded",
    "runtime": {
      "loadMs": null,
      "inferenceMs": null,
      "totalMs": 43624.3,
      "grade": "failed"
    },
    "failure": "No CPUExecutionProvider kernel was found for GatherBlockQuantized(1)."
  },
  {
    "precision": "q4f16",
    "device": "webgpu",
    "status": "pass",
    "quality": {
      "normalized": true,
      "finite": true,
      "expected": true,
      "nondegenerate": true,
      "grade": "typed-decision-pass"
    },
    "runtime": {
      "loadMs": 45219.2,
      "inferenceMs": 387.9,
      "totalMs": 47054.6,
      "grade": "good",
      "questionsPerSecond": 7.734
    }
  }
]

Something not quite right?

Tell us what happened with Open Jev DeBERTa 3. Loading, results, or the page itself—we’re here for all of it.

Attachments Optional

Up to 3 files · 50 MB each

Browser, OS, CPU/GPU, memory, screen, storage, and network details are included automatically. Model inputs, results, and browsing history are not collected.