EmbeddingGemma 2, text part only, MLX 4-bit

The text encoder of Google EmbeddingGemma 2, cut out of mlx-community/embeddinggemma-2-4bit (source revision 914f7f89142e33e77833254d9c9b90c3cef7303b). The picture, sound and video towers are not included: this copy turns text into a 768-number sentence vector and nothing else.

  • Weights: only the language_model.* tensors of the 4-bit conversion, unchanged (about 153 MB instead of 1.1 GB).
  • config.json is the original one; a loader that builds only the text model ignores its vision_config and audio_config.
  • Text vectors are identical to those of the full 4-bit conversion.

Prompts, as in the original model: a request is prefixed with task: search result | query: , a document with title: none | text: . Mean pooling over the tokens, then L2 normalisation.

Made for on-device search in the VoiceWork app (iPhone, MLX Swift).

Licence: Apache 2.0, as the original model. Attribution: Google.

Downloads last month
63
Safetensors
Model size
0.3B params
Tensor type
U32
ยท
BF16
ยท
MLX
Hardware compatibility
Log In to add your hardware

4-bit

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

Model tree for tarantila/embeddinggemma-2-text-4bit-mlx

Quantized
(146)
this model