EmbeddingGemma 2 for Core ML

AnyLab's conversion of Google's EmbeddingGemma 2 text encoder for Apple Silicon, targeting macOS 15 or later.

Output 768-dimensional, L2-normalized embeddings
Precision FP16 encoder; FP32 final normalization
Functions s128, s256, s512, s1024, s2048
Package 574.4 MB, with shared weights
License Apache-2.0

Download the model and tokenizer:

hf download anylab/embeddinggemma-2-coreml \
  --include 'EmbeddingGemma2Owned.mlpackage/**' --include 'tokenizer*.json' \
  --include LICENSE --include NOTICE --include 'licenses/QVAC-MIT.txt' \
  --local-dir embeddinggemma-2-coreml

Choose the smallest function that fits the input after prompts and special tokens. Load larger functions on demand and release them when idle; sharing weights on disk does not share their runtime allocations. First loading can take minutes. Core ML's cpuAndNeuralEngine setting still permits CPU work.

Validated on an M2 Air with 16 GB RAM: minimum cosine 0.999927 against 16 stock-model reference inputs. Full-app safety on 8 GB remains unmeasured.

The converter and detailed reports live in this repository. For deployments, pin an immutable HF revision and verify file hashes. This is an independent conversion, not an official Google release. See licence and attribution.

Downloads last month
80
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for anylab/embeddinggemma-2-coreml

Quantized
(146)
this model