---
license: apache-2.0
base_model:
- google/embeddinggemma-2
pipeline_tag: feature-extraction
tags:
- embeddinggemma
- embedding
- gewell
- bf16
- multimodal-embedding
- vision
- audio
- video
---

# EmbeddingGemma 2 — Gewell

This is the EmbeddingGemma 2 bundle used by
[Gewell](https://github.com/LeDissolution/gewell), a single-GPU inference
engine specialized for NVIDIA Blackwell (`sm_120`/`sm_120a`). It is a lossless
BF16 repack of `google/embeddinggemma-2` into Gewell's native embedding layout,
split into separately loadable text, vision and audio components. It is not a
Transformers, Sentence Transformers, vLLM, or llama.cpp checkpoint.

## Bundle contents

| File | Bytes | SHA-256 |
|---|---:|---|
| `text.safetensors` | 542,056,504 | `84bfdbbc9d5fe6dced9fd3b572329416fe2794c9c9e3a157346bc29b7c7ff361` |
| `vision.safetensors` | 335,543,360 | `2a146e1979e1fa9ac660aa871a2b405b92a0224b938ab4573dced85b1e43d600` |
| `audio.safetensors` | 611,314,696 | `ca511bf78d93dfc5b9d72bd61574e36087188a357e86f6fdbe65e034ae595017` |
| `config.json` | 4,455 | `b8f1e9931b57fbc054acdb445c41765d55b0074c58d145fa82839941ad1b5bb3` |
| `processor_config.json` | 1,788 | `168f6a08522f3ce5dea596d94d003af2fd691742d4f41fe1f9d8cce76bfbf69c` |
| `tokenizer.json` | 32,170,510 | `4d777ef5bdc1aa36227abdfb77c3e49e7b9c892d16e1b6bda41c393504828be4` |

The components hold 413 text, 211 vision/bridge and 752 audio/bridge tensors
(1,376 total). Tensor values are copied without conversion; native startup
validates the configuration, tokenizer, processor contract and tensor
inventory of every loaded component.

## Download and launch

Follow the [Gewell build instructions](https://github.com/LeDissolution/gewell/blob/main/docs/build.md),
then download this repository:

```bash
MODEL_DIR=/absolute/path/to/EmbeddingGemma-2-Gewell

hf download LeDissolution/EmbeddingGemma-2-Gewell --local-dir "$MODEL_DIR"

build/gewell serve-embeddings "$MODEL_DIR" --vision --audio
```

The server listens on `127.0.0.1:6311` and serves the OpenAI-compatible
`POST /v1/embeddings` endpoint under the model ID `google/embeddinggemma-2`:

```bash
curl -s http://127.0.0.1:6311/v1/embeddings \
  -H 'Content-Type: application/json' \
  -d '{"model":"google/embeddinggemma-2","input":["task: search result | query: What causes auroras?","title: none | text: Charged particles from the sun cause auroras."],"dimensions":768}'
```

Text is always loaded. `--vision` adds image and video input; `--audio` adds
audio input. Mixed image/video and audio requests need both. Explicit GPU
allocations at the default 8192-token capacity, excluding CUDA/cuBLAS/cuDNN
overhead:

| Options | Weight bytes | Scratch bytes |
|---|---:|---:|
| text only | 274,210,816 | 167,773,696 |
| `--vision` | 610,017,280 | 178,782,528 |
| `--vision --audio` | 1,223,507,968 | 1,028,441,648 |

The token-embedding table stays in host memory. Inputs support up to 8192
tokens including BOS/EOS and 128/256/512/768 output dimensions. Task prefixes
are not inserted automatically; supply them as described in the upstream
model card.

The engine was developed and tested on an RTX PRO 6000 Blackwell. See the
[model guide](https://github.com/LeDissolution/gewell/blob/main/docs/models.md),
[CLI reference](https://github.com/LeDissolution/gewell/blob/main/docs/cli.md)
and [HTTP API documentation](https://github.com/LeDissolution/gewell/blob/main/docs/http-api.md)
for request formats, media limits and serving options.

## Scope and provenance

This repository was created with

```bash
python3 tools/extract_embeddinggemma2.py --snapshot SNAPSHOT --vision --audio --output MODEL_DIR
```

from `google/embeddinggemma-2` at revision
`914f7f89142e33e77833254d9c9b90c3cef7303b`. Extraction is deterministic; you
can reproduce these files from the upstream snapshot instead of downloading
them.

Use is subject to the upstream
[EmbeddingGemma 2 license and terms](https://huggingface.co/google/embeddinggemma-2).