---
base_model: LCO-Embedding/LCO-Embedding-Omni-3B
library_name: llama.cpp
license: apache-2.0
language:
  - en
  - zh
  - multilingual
tags:
  - gguf
  - embedding
  - multimodal
  - qwen2
  - qwen2.5-omni
  - feature-extraction
  - text-embedding
  - image-embedding
  - audio-embedding
  - imatrix
pipeline_tag: feature-extraction
model_type: qwen2vl
quantized_by: marksverdhei
datasets:
  - MIEB
model_name: LCO-Embedding-Omni-3B-GGUF
model-index:
  - name: LCO-Embedding-Omni-3B-GGUF
    results:
      - task:
          type: feature-extraction
        dataset:
          name: MIEB-Lite
          type: MIEB
        metrics:
          - name: Embedding Dimensions
            type: embedding_dimensions
            value: 2048
---

# LCO-Embedding-Omni-3B-GGUF

GGUF quantizations of [LCO-Embedding/LCO-Embedding-Omni-3B](https://huggingface.co/LCO-Embedding/LCO-Embedding-Omni-3B) for use with [llama.cpp](https://github.com/ggml-org/llama.cpp).

Converted using [ht-llama.cpp](https://github.com/heiervang-technologies/ht-llama.cpp), a fork with added support for the `Qwen2_5OmniThinkerForConditionalGeneration` architecture.

## About the model

[LCO-Embedding-Omni-3B](https://huggingface.co/LCO-Embedding/LCO-Embedding-Omni-3B) is a multimodal embedding model based on the Thinker component of [Qwen 2.5 Omni](https://huggingface.co/Qwen/Qwen2.5-Omni-3B), fine-tuned with LoRA and contrastive learning to produce 2048-dimensional embeddings from text, images, audio, and video. Uses last-token pooling.

See [Scaling Language-Centric Omnimodal Representation Learning](https://arxiv.org/abs/2510.11693) (NeurIPS 2025) for details.

## Available files

### Standard quantizations

| File | Quant | Size | Description |
|------|-------|------|-------------|
| `LCO-Embedding-Omni-3B-BF16.gguf` | BF16 | -- | Full precision, no quality loss |
| `LCO-Embedding-Omni-3B-Q8_0.gguf` | Q8_0 | -- | Near-lossless quantization |
| `LCO-Embedding-Omni-3B-Q4_K_M.gguf` | Q4_K_M | -- | Good balance of quality and size |
| `LCO-Embedding-Omni-3B-Q3_K_M.gguf` | Q3_K_M | -- | Smaller, some quality loss |
| `LCO-Embedding-Omni-3B-Q2_K.gguf` | Q2_K | -- | Smallest, more quality loss |

### Importance matrix (imatrix) quantizations

Quantized with an importance matrix computed from WikiText-2 calibration data for improved quality at low bit widths.

| File | Quant | Size | Description |
|------|-------|------|-------------|
| `LCO-Embedding-Omni-3B-IQ4_XS.gguf` | IQ4_XS | -- | 4.25 bpw, imatrix-optimized |
| `LCO-Embedding-Omni-3B-IQ3_M.gguf` | IQ3_M | -- | 3.66 bpw, imatrix-optimized |
| `LCO-Embedding-Omni-3B-IQ3_XS.gguf` | IQ3_XS | -- | 3.3 bpw, imatrix-optimized |
| `LCO-Embedding-Omni-3B-IQ2_M.gguf` | IQ2_M | -- | 2.7 bpw, imatrix-optimized |

### Multimodal projection

| File | Quant | Size | Description |
|------|-------|------|-------------|
| `mmproj-LCO-Embedding-Omni-3b-F16.gguf` | F16 | -- | Vision + audio projection (required for multimodal) |

For text-only embedding, you only need one of the text model GGUFs. For multimodal (image/audio/video), you also need the `mmproj` file.

## Quantization quality

Measured on 8 diverse text sentences (2048-dim embeddings). BF16 is the reference.

### Embedding quality vs BF16

*Results will be added after quantization testing.*

### pgvector retrieval quality (query with quant, corpus in BF16)

*Results will be added after quantization testing.*

## Usage

### Build llama.cpp

```bash
git clone https://github.com/heiervang-technologies/ht-llama.cpp
cd ht-llama.cpp
cmake -B build
cmake --build build --target llama-embedding llama-server -j$(nproc)
```

### Text embeddings (CLI)

```bash
./build/bin/llama-embedding \
  -m LCO-Embedding-Omni-3B-Q8_0.gguf \
  --pooling last \
  -p "Your text here"
```

### Text embeddings (server)

```bash
./build/bin/llama-server \
  -m LCO-Embedding-Omni-3B-Q8_0.gguf \
  --embedding --pooling last

curl -s http://localhost:8080/embeddings \
  -d '{"content": "Your text here"}'
```

### Multimodal embeddings (vision + audio)

Requires the mmproj file:

```bash
./build/bin/llama-server \
  -m LCO-Embedding-Omni-3B-Q8_0.gguf \
  --mmproj mmproj-LCO-Embedding-Omni-3b-F16.gguf \
  --embedding --pooling last
```

```bash
# Image embedding (base64-encoded image)
curl -s http://localhost:8080/embeddings \
  -d '{"content": [{"prompt_string": "<__media__>", "multimodal_data": ["<base64-image-data>"]}]}'

# Audio embedding (base64-encoded WAV)
curl -s http://localhost:8080/embeddings \
  -d '{"content": [{"prompt_string": "<__media__>", "multimodal_data": ["<base64-audio-data>"]}]}'
```

### JSON output (for programmatic use)

```bash
./build/bin/llama-embedding \
  -m LCO-Embedding-Omni-3B-Q8_0.gguf \
  --pooling last \
  --embd-output-format json \
  -p "Your text here"
```

## Notes

- This is a quantization of [LCO-Embedding/LCO-Embedding-Omni-3B](https://huggingface.co/LCO-Embedding/LCO-Embedding-Omni-3B) -- see the original model card for benchmarks, training details, and licensing
- The `--pooling last` flag is required -- this model uses last-token pooling, not mean pooling
- Embedding dimensions: 2048
- Contributions and bug reports welcome at [ht-llama.cpp](https://github.com/heiervang-technologies/ht-llama.cpp/issues)

## Citations

### LCO-Embedding

```bibtex
@article{xiao2025scaling,
  title={Scaling Language-Centric Omnimodal Representation Learning},
  author={Xiao, Chenghao and Chan, Hou Pong and Zhang, Hao and Xu, Weiwen and Aljunied, Mahani and Rong, Yu},
  journal={arXiv preprint arXiv:2510.11693},
  year={2025}
}
```

### Qwen 2.5 Omni

```bibtex
@article{Qwen2.5-Omni,
  title={Qwen2.5-Omni Technical Report},
  author={Jin Xu and Zhifang Guo and Jinzheng He and Hangrui Hu and Ting He and Shuai Bai and Keqin Chen and Jialin Wang and Yang Fan and Kai Dang and Bin Zhang and Xiong Wang and Yunfei Chu and Junyang Lin},
  journal={arXiv preprint arXiv:2503.20215},
  year={2025}
}
```