Models · Qwen · Qwen3-ASR 1.7B
Sabich-Qwen3-ASR-1.7B desktop GPU
Sabich's training data on Qwen3-ASR 1.7B, an ASR model with an LLM decoder. The most accurate Sabich, for computers with a GPU: about 5 GB of VRAM in bf16, or about 2 GB as a GGUF.
At a glance
| Base model | Qwen/Qwen3-ASR-1.7B (Qwen) |
| Current version | Sabich-Qwen3-ASR-1.7B v1.0, released 2026-10-21 |
| Parameters | 2.0B (Hub safetensors count; the base is named 1.7B) |
| Evaluation | 5.4% code-switching WER (stock Qwen3-ASR 10.0%) on the held-out split at revision cf84f7e4 Results on this page · Model card |
| Download | every format and quant (Hugging Face) |
| Collection | Hebrish/sabich on Hugging Face |
| Licence | Apache-2.0 (base model) · use governed by the terms of use |
| Used by | Your own tools (Transformers, llama.cpp); not in the Hebrish apps yet |
| DOIs | 10.57967/hf/11173 |
Formats and quants
| Repository | Format and files | Version | For | DOI |
|---|---|---|---|---|
Hebrish/Sabich-Qwen3-ASR-1.7B main | Transformers (safetensors, bf16) model.safetensors 4.08 GB, tokenizer, training config · access request · files | v1.0 | qwen-asr / Transformers on PyTorch CUDA or ROCm, about 5 GB of VRAM | 10.57967/hf/11173 |
Hebrish/Sabich-Qwen3-ASR-1.7B-GGUF | GGUF llama.cpp: Q8_0 1.83 GB + mmproj 356 MB (also bf16, Q4_K_M) · transcribe.cpp: Q4_K_M 1.32 GB, Q8_0 2.19 GB · access request · files | v1.0 | llama.cpp (HIP, Vulkan, CUDA); transcribe.cpp (Offline Voice Input) | none yet |
Hugging Face issues a DOI per repository, so each format can be cited on its own. Repositories, sizes and DOIs checked on the Hub on 2026-10-21.
Versions
| Version | Released | Run | Where | Notes |
|---|---|---|---|---|
| Sabich-Qwen3-ASR-1.7B v1.0 current | 2026-10-21 | q1-all | Sabich-Qwen3-ASR-1.7B | Sabich's data (blend all, revision e918e381): 592 min of the speaker's takes (383 min code-switching) plus 66 min LibriSpeech replay; 3 epochs on one H100, epoch 3 kept |
About Sabich-Qwen3-ASR-1.7B
Sabich-Qwen3-ASR-1.7B v1.0 (2026-10-21) is Sabich, the everything model (code-switching plus tech vocabulary, brand names, numbers and the other task recordings), trained on a larger base. Each line keeps its food name; a variant on another base adds the base: Sabich-<Base>-<Size>. It is not in the Hebrish apps yet: run it in your own tools.
Against stock Qwen3-ASR 1.7B, on the held-out test split of the speaker's recordings at dataset revision cf84f7e4 (term-disjoint from training; normalised WER):
| Stock Qwen3-ASR-1.7B | Sabich-Qwen3-ASR-1.7B | |
|---|---|---|
| English-Hebrew code-switching, WER | 10.0% | 5.4% |
| Hebrew terms never seen in training, recall | 32% | 50% |
| Hebrew terms seen in training, recall | 40% | 73% |
| Speaker's own English, WER | 4.6% | 2.5% |
| LibriSpeech test-clean (200 utterances), WER | 1.2% | 1.35% |
| Long-form (about 3-minute joined clips), WER | 12.0% | 8.7% |
| Numbers written in digits | 2/8 | 8/8 |
Run q1-all: Sabich's data at revision e918e381, 592 min of the speaker's takes (383 min code-switching) plus 10% LibriSpeech-PC replay, 658 min in all. The official Qwen3-ASR full fine-tuning script, 3 epochs, one H100, 15 minutes; epoch 3 kept. These numbers are on a different test-set revision from Sabich's, so the two have not been compared head to head yet. Long-form audio is its weakest area, and about half of unseen Hebrew words are still missed or misspelled.
Run it
Transformers / PyTorch, on CUDA or ROCm, about 5 GB of VRAM in bf16. Always pass language="English", the prefix it was trained with:
import torch
from qwen_asr import Qwen3ASRModel # pip install qwen-asr
model = Qwen3ASRModel.from_pretrained(
"Hebrish/Sabich-Qwen3-ASR-1.7B", dtype=torch.bfloat16, device_map="cuda:0")
print(model.transcribe(audio="note.wav", language="English")[0].text)- GGUF with llama.cpp: HIP or Vulkan on AMD, CUDA on NVIDIA. The Q8_0 file is about 2 GB, so it fits a 4 GB card. Files:
Hebrish/Sabich-Qwen3-ASR-1.7B-GGUF. - On Android, a transcribe.cpp Q4_K_M GGUF runs in Offline Voice Input, but slowly. For phones, Sabich is the better fit.
- Output is English, cased and punctuated, with Hebrew words in Latin letters, never Hebrew script. Numbers come out in digits.
Cite
Cite Sabich-Qwen3-ASR-1.7B with its DOI, 10.57967/hf/11173 (Hebrish/Sabich-Qwen3-ASR-1.7B). To cite one format, use that repository's DOI from the table above.
@misc{rosehill_sabich_qwen3_2026,
author = {Daniel Rosehill},
title = {Sabich-Qwen3-ASR-1.7B},
year = {2026},
publisher = {Hugging Face},
url = {https://huggingface.co/Hebrish/Sabich-Qwen3-ASR-1.7B},
doi = {10.57967/hf/11173}
}Request access
Request access on Hugging Face first. The Sabich-Qwen3-ASR-1.7B repositories are gated: open the repository, sign in, and fill in the short request form there. Requests are reviewed by hand. For non-commercial use, the attestation form covers what is asked.
No answer after a few days? Request access here instead, or write to access-requests@hebrish.app. Commercial use, including telephony, needs a commercial licence.
Which model should I use? · All Sabich variants · All Hebrish models