gety-embed-v0
Fine-tuned from
intfloat/multilingual-e5-small using open-source and proprietary synthetic data, optimized for local search scenarios.
ONNX
| File | Quantization | Size |
|---|
onnx/model_uint8.onnx | UINT8 Dynamic | 112 MB |
ONNX CUDA EP
| File | Quantization | Size |
|---|
onnx_cuda/model_fp16.onnx | FP16 | 224 MB |
Inputs input_ids + attention_mask; masked mean pooling + L2 normalization
baked into the graph, output is the normalized 384-dim embedding.
Core ML
| File | Quantization | Size |
|---|
coreml/model_fp16.mlpackage | FP16 | 224 MB |
Single input_ids input (dynamic 1-512 tokens, no padding), pooling and
normalization baked in. macOS 13+, ANE accelerated.
OpenVINO (Intel NPU)
| File | Quantization | Size |
|---|
openvino/model.xml + openvino/model.bin | INT8 weight-only (NNCF), FP16 activations | 114 MB |
Single-input OpenVINO IR for static-shape NPU compilation
(reshape([1, L]) + compile, batch=1, sequence buckets e.g. 32/64/128/256/512):
- Input:
input_ids [batch, seq] (int64), padded to the bucket length with
<pad> (token id 1 — not config.json's pad_token_id=0, which is <s>)
- The attention mask is computed in-graph as
input_ids != 1
- Masked mean pooling + L2 normalization baked in
- Output:
embeddings [batch, 384] (float32)
Selected over FP16 / static-PTQ INT8 / INT4 on NPU hardware (Intel AI Boost,
OpenVINO 2026.2.1): equal latency to all other variants (NPU latency is
shape-bound), best accuracy (min cosine 0.9998 vs PyTorch reference; static
PTQ drops to 0.9872 on short queries), and half the size of FP16.
