google/embeddinggemma-2

Feature Extraction • 0.7B • Updated 14 days ago • 172k • 2.75k

Note The original bf16 model from Google that every build here is quantized from (1489 MB).

srv-sngh/embeddinggemma-2-mlx-mixed-mxfp8

Feature Extraction • 0.7B • Updated 12 days ago • 99

Note 1007 MB. mxfp8 on the embedding table, MLP, vision and audio weights; text layers and attention kept bf16.

srv-sngh/embeddinggemma-2-mlx-8bit

Feature Extraction • 0.7B • Updated 12 days ago • 120

Note 806 MB. Affine 8-bit, group 64, every quantizable tensor.

srv-sngh/embeddinggemma-2-mlx-6bit

Feature Extraction • 0.7B • Updated 12 days ago • 77

Note 624 MB. Affine 6-bit, group 64, every quantizable tensor.

srv-sngh/embeddinggemma-2-mlx-6bit-dyn

Feature Extraction • 0.7B • Updated 12 days ago • 61

Note 657 MB. 6-bit with text layers at 8-bit. Not recommended over plain 6bit.

srv-sngh/embeddinggemma-2-mlx-5bit

Feature Extraction • 0.7B • Updated 12 days ago • 68

Note 533 MB. Affine 5-bit, group 64, every quantizable tensor.

srv-sngh/embeddinggemma-2-mlx-4bit-dyn

Feature Extraction • 0.7B • Updated 12 days ago • 68

Note 473 MB. 4-bit with text attention, per-layer inputs, projection and last two text layers at 8-bit.

srv-sngh/embeddinggemma-2-mlx-4bit-dyn-text

Feature Extraction • 0.7B • Updated 12 days ago • 118

Note 507 MB. 4-bit with all 24 text layers and the text projection at 8-bit; vision/audio stay 4-bit.

srv-sngh/embeddinggemma-2-mlx-4bit-g32

Feature Extraction • 0.7B • Updated 12 days ago • 117

Note 488 MB. Affine 4-bit, group 32, every quantizable tensor.

srv-sngh/embeddinggemma-2-mlx-nvfp4

Feature Extraction • 0.7B • Updated 12 days ago • 212

Note 442 MB. nvfp4 (4-bit float, group 16, fp8 scale), the smallest build.