| File | Size | SHA-256 |
|---|---|---|
Qwen3-ASR-1.7B-Q4_K_M.gguf | 1.28 GB | f53b45df729c90a182612b759227daa9bb72139b012a87e558dd6c204d486ae9 |
Qwen3-ASR-1.7B-Q5_K_M.gguf | 1.26 GB | 33ca947c1a39f80f5fe0774324924dd11f603ef4c08b50aa04ead3133787e54c |
Qwen3-ASR-1.7B-Q8_0.gguf | 2.17 GB | 58e22d0532d4eacaf034cfac17a6fed159f37c41390c710186783be439d1fc57 |
mmproj-Qwen3-ASR-1.7B-Q8_0.gguf | 356 MB | 46c1d533af3f354ceb37ce855dbceff7da7fa7cf1e6a523df3b13440bd164c0d |
Q8_0 and mmproj are unmodified copies from ggml-org/Qwen3-ASR-1.7B-GGUF (identical SHA-256). Q4_K_M is requantized from that Q8_0.Q5_K_M is converted from the original bf16 weights and quantized directly. The tied embedding matrix, which the converter writes twice, is stored once: output.weight is dropped and token_embd.weight kept at Q6_K, which runtimes substitute for it — hence the file is smaller than Q4_K_M.