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.

EmbeddingGemma

Compare meaning, explore text vectors, and build semantic search on your device.

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

Convert multilingual text into vectors for semantic search and similarity matching. Returns normalized 768-number vectors, optionally shortened to 512, 256, or 128 values; queries and documents must use the expected formatting.

Created by
Google
Model released
Sep 4, 2025
Source repository created
Jul 17, 2025
Model weights · Q4
197.2 MB

Best for

  • Semantic search and retrieval across more than 100 trained languages.
  • Text similarity, clustering, classification, question answering, fact-verification retrieval, and code retrieval.

Keep in mind

  • Retrieval quality depends on using the documented query/document prefixes and varies by language, domain, ambiguity, and input quality.
  • Inputs longer than 2,048 tokens are truncated by the model.
  • Embedding similarity is not factual verification and can reproduce training-data gaps or bias.
  • FP16 activations and derivatives are explicitly unsupported upstream; quantization and Matryoshka truncation can alter ranking margins.

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
google/embeddinggemma-300m
Release verification
Verified
Release evidence
Original Google release announcement identifies EmbeddingGemma 300M.
Source updated
Sep 25, 2025
Browser model files
onnx-community/embeddinggemma-300m-ONNX
Files updated
Sep 4, 2025
Artifact revision
5090578d9565bb06545b4552f76e6bc2c93e4a66

Input

TEXT → VECTORS

Sample result

EmbeddingGemma 300M

Settings

Matryoshka output dimension; truncated vectors are re-normalized. Lower dimensions reduce storage and similarity-comparison work, but can reduce retrieval quality. Higher dimensions retain more information at greater storage cost. Vectors compared together must use the same dimension.

Output

Explore what your text means

Generate vectors from one or more texts. Add multiple texts to compare their meaning.

Benchmark evidence

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

EmbeddingGemma 300M

12 configurations tested
Passed8
Failed4
TestedSep 12, 2026
RuntimeTransformers.js 4.2.0

Quality criteria

Finite normalized vectors with the correct dimensions, distinct rows, and the relevant-document cosine exceeding the unrelated-document cosine by at least 0.1.

PrecisionDeviceResult
Q4GPU · WebGPUAcceleratedPassedVery fast
Q8GPU · WebGPUAcceleratedPassedFast
NO_GATHER_Q4GPU · WebGPUAcceleratedPassedFast
FP32GPU · WebGPUAcceleratedPassedFast
FP32CPU · WASMPassedFast
FP16CPU · WASMPassedFast
Q8CPU · WASMPassedFast
NO_GATHER_Q4CPU · WASMPassedModerate
FP16GPU · WebGPUFailed—
Q4CPU · WASMFailed—
Q4F16CPU · WASMFailed—
Q4F16GPU · WebGPUFailed—
Test setup & sample details
Quality gate
Finite normalized vectors with the correct dimensions, distinct rows, and the relevant-document cosine exceeding the unrelated-document cosine by at least 0.1.
Tested
2026-09-12T14:42:07.406Z
Environment
Not recorded
Runtime
Transformers.js 4.2.0
Artifact revision
5090578d9565bb06545b4552f76e6bc2c93e4a66
Fixture
{
  "file": "embeddinggemma-300m.json",
  "url": "https://f005.backblazeb2.com/file/hubters-webai-public/assets/benchmarks/embeddinggemma-300m.json",
  "size": 634,
  "sha1": "9a1d261338de718b559f5544e81e488477672779"
}
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": {
      "margin": 0.6141
    },
    "runtime": {
      "loadMs": 140152.69999999553,
      "inferenceMs": 320.6000000014901,
      "totalMs": 141775.79999999702,
      "embeddingsPerSecond": 9.3575,
      "grade": "slow"
    }
  },
  {
    "precision": "fp32",
    "device": "webgpu",
    "status": "pass",
    "quality": {
      "margin": 0.6141
    },
    "runtime": {
      "loadMs": 159880.8999999985,
      "inferenceMs": 251.30000000447035,
      "totalMs": 161532.29999999702,
      "embeddingsPerSecond": 11.9379,
      "grade": "good"
    }
  },
  {
    "precision": "fp16",
    "device": "wasm",
    "status": "pass",
    "quality": {
      "margin": 0.6141
    },
    "runtime": {
      "loadMs": 85957.60000000149,
      "inferenceMs": 337.70000000298023,
      "totalMs": 87670.39999999851,
      "embeddingsPerSecond": 8.8836,
      "grade": "slow"
    }
  },
  {
    "precision": "fp16",
    "device": "webgpu",
    "status": "fail",
    "quality": "Not embedded",
    "runtime": "Not embedded",
    "failure": "WebGPU LayerNorm shader validation failed during inference."
  },
  {
    "precision": "q8",
    "device": "wasm",
    "status": "pass",
    "quality": {
      "margin": 0.6085
    },
    "runtime": {
      "loadMs": 49854.39999999851,
      "inferenceMs": 445.6000000014901,
      "totalMs": 51707.5,
      "embeddingsPerSecond": 6.7325,
      "grade": "slow"
    }
  },
  {
    "precision": "q8",
    "device": "webgpu",
    "status": "pass",
    "quality": {
      "margin": 0.6085
    },
    "runtime": {
      "loadMs": 46093.30000000447,
      "inferenceMs": 174.89999999850988,
      "totalMs": 47672.39999999851,
      "embeddingsPerSecond": 17.1527,
      "grade": "good"
    }
  },
  {
    "precision": "q4",
    "device": "wasm",
    "status": "fail",
    "quality": "Not embedded",
    "runtime": "Not embedded",
    "failure": "WASM has no implementation for GatherBlockQuantized."
  },
  {
    "precision": "q4",
    "device": "webgpu",
    "status": "pass",
    "quality": {
      "margin": 0.5983
    },
    "runtime": {
      "loadMs": 36447.19999999553,
      "inferenceMs": 80.60000000149012,
      "totalMs": 37903.39999999851,
      "embeddingsPerSecond": 37.2208,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4f16",
    "device": "wasm",
    "status": "fail",
    "quality": "Not embedded",
    "runtime": "Not embedded",
    "failure": "WASM has no GatherBlockQuantized kernel."
  },
  {
    "precision": "q4f16",
    "device": "webgpu",
    "status": "fail",
    "quality": "Not embedded",
    "runtime": "Not embedded",
    "failure": "WebGPU LayerNorm shader validation failed during inference."
  },
  {
    "precision": "no_gather_q4",
    "device": "wasm",
    "status": "pass",
    "quality": {
      "margin": 0.5969
    },
    "runtime": {
      "loadMs": 28191.29999999702,
      "inferenceMs": 1558.5,
      "totalMs": 30994,
      "embeddingsPerSecond": 1.9249,
      "grade": "very-slow"
    }
  },
  {
    "precision": "no_gather_q4",
    "device": "webgpu",
    "status": "pass",
    "quality": {
      "margin": 0.5969
    },
    "runtime": {
      "loadMs": 26292,
      "inferenceMs": 188,
      "totalMs": 27748.89999999851,
      "embeddingsPerSecond": 15.9574,
      "grade": "good"
    }
  }
]

Something not quite right?

Tell us what happened with EmbeddingGemma. 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.