ThakiCloud/Qwen3-Embedding-0.6B-MultiCal-M1a-GPTQ
Feature Extraction • 0.6B • Updated • 89
Low-bit quantized multilingual retrieval embedders, shipped in a format a standard loader runs: mixed 3/4-bit packed artifacts plus a GGUF re-encode.
Note Qwen3-Embedding-0.6B at mixed precision (4-bit g64 layers 0-6, 3-bit g32 layers 7-27, 3-bit g64 embedding table); packed file 295,391,600 B. Ships the runnable Q4_1 GGUF re-encode (378,522,208 B): 197/197 tensors bit-identical on read-back, token ids match HF 9/9, min cosine 0.999020 vs our fp32 reference, 3/3 top-1 on a smoke set. nDCG was measured on the packed artifact, not through llama.cpp. Research Preview.