ThakiCloud/eg2-text-q4

Feature Extraction • 0.1B • Updated 6 days ago • 88

Note 4-bit text tower, 152,669,232 B of weights. Keeps 0.936 of the BF16 top-10 neighborhood (BF16 = 1.000); margin 0.1712 vs BF16 0.1817. Internally measured on a private frozen set; serving not validated.

ThakiCloud/eg2-text-hybrid-129m

Feature Extraction • 0.1B • Updated 6 days ago • 70

Note Mixed-precision text tower (GPTQ 4-bit transformer, 2-bit vocabulary with 30% of rows at 4-bit), 129,532,796 B — 85% of eg2-text-q4. Beats uniform 3-bit on every held-out slice; within 25% of the 3→4-bit gap of the Q4 recipe. Internally measured; serving not validated.

ThakiCloud/eg2-text-hybrid-119m

Feature Extraction • 0.1B • Updated 6 days ago • 73

Note Size-oriented mixed-precision text tower (GPTQ 4-bit transformer, 2-bit vocabulary, no row rescue), 119,152,312 B. At the uniform-3-bit byte budget it wins all 8 held-out cells over uniform 3-bit, and it gives up most of the 4-bit advantage on multilingual and rare-identifier retrieval. Internally measured; serving not validated.

Note Why this line allocates bits the way it does. The embedding table never emerges as the dominant isolated protection priority in any of the four families measured, and at INT4/g16 the spread between modules is too small to allocate against. Measures EmbeddingGemma-300M, not EmbeddingGemma 2.

Note The label-free drift signal this line's mixed-precision allocation descends from, measured on five development embedders including EmbeddingGemma-300M. EmbeddingGemma 2 itself is not in that paper's grid.

ThakiCloud/eg2-hybrid-w4-media

Feature Extraction • 0.5B • Updated 15 days ago • 63

Note Multimodal bundle: BF16 text tower, 4-bit vision/audio towers. Model files 871,819,777 B (57% of upstream BF16; repo adds the card, Korean card and LICENSE). Image, audio, video retrieval met pre-registered non-inferiority. Load via the bundled eg2_quant.py (dequantizes at load). For lower GPU memory on B200 use the eg2-l2v1 runtime (github.com/ThakiCloud/eg2-l2v1): 0.58x resident memory vs BF16, bit-identical embeddings, not faster.

ThakiCloud/eg2-mixed-bit-550mb

Feature Extraction • 0.5B • Updated 7 days ago • 54

Note INDEX side. 549,949,165 B (36% of upstream BF16), 4/6/8 bits per matrix allocated by measured sensitivity. Sealed evaluation (run once): image and video retrieval non-inferior to BF16 (IMAGE_VIDEO_GO); audio and text-only not evaluated, so NO_MULTIMODAL_GO. H100 to B200 rebuild reproduces the decoded tensors. Not shown to beat the same-size uniform control.

ThakiCloud/eg2-w6-text-w4-media

Feature Extraction • 0.5B • Updated 7 days ago • 58

Note INDEX side. Uniform-precision control for eg2-mixed-bit-550mb: 6-bit text, 4-bit vision/audio, 550,244,021 B, same quantizer and group size. Sealed evaluation (run once): image and video retrieval non-inferior to BF16 (IMAGE_VIDEO_GO); audio and text-only not evaluated, so NO_MULTIMODAL_GO. Two B200 builds reproduce the decoded tensors (not tested across GPU types).

ThakiCloud/eg2-tiny-query-a1

Feature Extraction • 15.3M • Updated 7 days ago • 38

Note QUERY side (index with google/embeddinggemma-2). 17.6 MB INT8 Korean/English query encoder (TinyCLIP Text-3M + 768 projection) that searches existing EmbeddingGemma 2 image vectors without re-indexing. English app-screen R@10 kept vs its English-only predecessor; Korean charts improve only when fused with OCR + BM25 (alone it trails BM25). Single seed; platform screens not tested.

ThakiCloud/eg2-tiny-query-commerce-ko

Feature Extraction • Updated 7 days ago

Note QUERY side, recommended version K4 (main). Two 15.4 MB INT8 query encoders (Korean K4 + English E0, Hangul routing) that search an EmbeddingGemma 2 product-image index without re-indexing; build that index with google/embeddinggemma-2. Korean product-name R@10 0.368 -> 0.711 vs E0 (development data). Use BM25 for exact product names; the model is for descriptive queries. Usage guide: https://thakicloud.com/tech-blog/en/tutorials/eg2-edge-usage-guide/

Note Learned PTQ calibration at 4-bit on e5-base-v2, BGE-M3 and EmbeddingGemma-300M: the update dose decides gain or collapse (+1.47 to -71.53 NDCG@10 on EmbeddingGemma-300M). A validation-gated accept rule made 0 harmful picks in 30 decisions at 1.52 pp mean regret. No checkpoint here uses learned calibration, and EmbeddingGemma 2 is not in the grid.