| # Provenance — majentik/Qwen3-ASR-1.7B-MLX-8bit | |
| - Upstream repo: `Qwen/Qwen3-ASR-1.7B-hf` | |
| - Upstream revision: `057a3b044fcd31c433e7971ab40d68d20e7eae6d` | |
| - Quantization: bits=8, mode=affine, group_size=64 | |
| - Command: `python -m pipelines.mlx_direct_quantize --model qwen3-asr-1.7b --base-dir /private/tmp/mlx-direct-release/qwen3-asr-1.7b/base --out-dir /tmp/mlx-direct-release/qwen3-asr-1.7b/8bit --bits 8 --mode affine --group-size 64` | |
| - Input bytes (BF16 base): 4087645152 | |
| - Output bytes (this variant): 2474847246 | |
| - Toolchain: mlx 0.31.2, huggingface_hub 0.36.2 | |
| - Generated: 2026-07-03T14:23:01.955115+00:00 | |
| Audio-inference smoke was performed post-upload on 2026-07-03; this variant passed its CER gate — see the README's Audio inference smoke section. General runtime support remains pending upstream mlx-lm support for the `qwen3_asr` audio-text architecture (the Hugging Face runtime requires transformers 5.13+ — `qwen3_asr` merged after the 5.12.1 release). | |