Skip to content
Other

Qwen3-ASR-1.7B (FP8 W8A8)

fast quantized model for speech recognition

2 min readupdated

This model is an optimized quantized version (FP8 W8A8) of the modern Qwen3-ASR-1.7B speech recognition and acoustic system. It is designed for high-quality transcription of audio files, songs, and conversational speech, along with on-the-fly language identification for 30 languages and 22 dialects (including Russian). Thanks to targeted quantization of the decoder linear layers, the model size is reduced by more than a third while maintaining high accuracy. The model supports running via convenient transformers and vLLM (high-speed generation server) backends for efficient inference.

Key features

  • FP8 W8A8 quantization: decoder linear layer weights are compressed to the FP8 format, while the audio encoder and projector remain in BF16 to preserve acoustic accuracy.
  • Support for 52 languages and dialects: confident recognition of speech, accents, and songs in Russian, English, Chinese, and other languages.
  • Unified inference pipeline: support for batch processing, streaming mode, and standalone transcription of long audio recordings.
  • vLLM compatibility: high processing speed and the ability to deploy via a standard OpenAI-compatible server.

What you can do

  • Run a local transcription service: deploy a speech recognition server using vLLM to process audio streams in real time.
  • Transcribe long podcasts and videos: automate the transcription of multi-hour audio recordings with high speed and minimal VRAM requirements.
  • Integrate into voice bots: use the model as a backend for recognizing user speech in multimodal applications.

Example

import torch
from qwen_asr import Qwen3ASRModel

model = Qwen3ASRModel.from_pretrained(
    "lly0571/Qwen3-ASR-1.7B-FP8-W8A8",
    dtype=torch.bfloat16,
    device_map="cuda:0"
)

results = model.transcribe(
    audio="[https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-ASR-Repo/asr_en.wav](https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-ASR-Repo/asr_en.wav)",
    language=None
)
print(results[0].text)
Read moreDetails
Weight size: 2.48 GB (-34.6% compared to the baseline BF16 version).
Recognition quality (WER on LibriSpeech test-clean): 1.64% (virtually no loss against 1.62% for the original).
Optimization: 196 decoder linear layers quantized using LLM Compressor tools.
Examples
Discuss Telegram 0 21

Will it run?

Runs in F8_E4M3

Fits without quantisation.

  • F8_E4M3recommendedweights 2.5 GB · ready files3.2 GB
  • 4-bitweights 1.1 GB1.7 GB

Tip: go 8-bit.F8_E4M3 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: F8_E4M3 — from RTX 2060

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # lly0571/Qwen3-ASR-1.7B-FP8-W8A8
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download lly0571/Qwen3-ASR-1.7B-FP8-W8A8 --local-dir ~/models/Qwen3-ASR-1.7B-FP8-W8A8
    # The model is in ~/models. How to run it is in the model card (link above).
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

New posts about versions of Qwen3-ASR-1.7B-FP8-W8A8 (quants, finetunes, LoRAs) will show up under the bell.

No known relatives of this model yet. All derivatives on Hugging Face

About the model

lly0571/Qwen3-ASR-1.7B-FP8-W8A8
Author
lly0571
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

210

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments