cstr/qwen3-asr-1.7b-gguf overview

{
  "metadata": {},
  "card_data": {
    "license": "apache-2.0",
    "base_model": "Qwen/Qwen3-ASR-1.7B",
    "language": [
      "zh",
      "en",
      "yue",
      "ar",
      "de",
      "fr",
      "es",
      "pt",
      "id",
      "it",
      "ko",
      "ru",
      "th",
      "vi",
      "ja",
      "tr",
      "hi",
      "ms",
      "nl",
      "sv",
      "da",
      "fi",
      "pl",
      "cs",
      "fil",
      "fa",
      "el",
      "ro",
      "hu",
      "mk"
    ],
    "tags": [
      "asr",
      "speech-to-text",
      "audio",
      "transcription",
      "gguf",
      "ggml",
      "speech-llm",
      "qwen3"
    ],
    "pipeline_tag": "automatic-speech-recognition",
    "library_name": "ggml",
    "frontmatter": {
      "license": "apache-2.0",
      "base_model": "Qwen/Qwen3-ASR-1.7B",
      "language": [
        "zh",
        "en",
        "yue",
        "ar",
        "de",
        "fr",
        "es",
        "pt",
        "id",
        "it",
        "ko",
        "ru",
        "th",
        "vi",
        "ja",
        "tr",
        "hi",
        "ms",
        "nl",
        "sv",
        "da",
        "fi",
        "pl",
        "cs",
        "fil",
        "fa",
        "el",
        "ro",
        "hu",
        "mk"
      ],
      "tags": [
        "asr",
        "speech-to-text",
        "audio",
        "transcription",
        "gguf",
        "ggml",
        "speech-llm",
        "qwen3"
      ],
      "pipeline_tag": "automatic-speech-recognition",
      "library_name": "ggml"
    },
    "hero_image_url": "",
    "summary": "GGUF conversions of Qwen/Qwen3-ASR-1.7B, the larger of the two Qwen3-ASR speech-LLMs (Whisper-style audio encoder + 1.7B Qwen3 LLM head). Runs through CrispASR's --backend qwen3 path with no architecture changes — the same C++ qwen3 backend that already supported the 0.6B variant loads the 1.7B directly because the converter and the runtime both read sizes from GGUF metadata rather than hardcoding them.",
    "quick_links": [],
    "benchmark_table_html": "",
    "readme_markdown": "---\nlicense: apache-2.0\nbase_model: Qwen/Qwen3-ASR-1.7B\nlanguage:\n  - zh\n  - en\n  - yue\n  - ar\n  - de\n  - fr\n  - es\n  - pt\n  - id\n  - it\n  - ko\n  - ru\n  - th\n  - vi\n  - ja\n  - tr\n  - hi\n  - ms\n  - nl\n  - sv\n  - da\n  - fi\n  - pl\n  - cs\n  - fil\n  - fa\n  - el\n  - ro\n  - hu\n  - mk\ntags:\n  - asr\n  - speech-to-text\n  - audio\n  - transcription\n  - gguf\n  - ggml\n  - speech-llm\n  - qwen3\npipeline_tag: automatic-speech-recognition\nlibrary_name: ggml\n---\n\n# Qwen3-ASR-1.7B — GGUF (CrispASR)\n\nGGUF conversions of [`Qwen/Qwen3-ASR-1.7B`](https://huggingface.co/Qwen/Qwen3-ASR-1.7B), the larger of the two Qwen3-ASR speech-LLMs (Whisper-style audio encoder + 1.7B Qwen3 LLM head). Runs through [CrispASR](https://github.com/CrispStrobe/CrispASR)'s `--backend qwen3` path with no architecture changes — the same C++ qwen3 backend that already supported the 0.6B variant loads the 1.7B directly because the converter and the runtime both read sizes from GGUF metadata rather than hardcoding them.\n\n## What's in the box\n\n| File | Size | Quantization | Notes |\n|---|---|---|---|\n| `qwen3-asr-1.7b-f16.gguf` | 4.71 GB | F16 | Reference precision; matches PyTorch bfloat16 within float-noise tolerance |\n| `qwen3-asr-1.7b-q8_0.gguf` | 2.51 GB | Q8_0 | Effectively lossless; same transcript on jfk.wav as F16 |\n| `qwen3-asr-1.7b-q4_k.gguf` | 1.33 GB | Q4_K | 3.5× compressed; minor punctuation differences on tested clips |\n\nAll three contain:\n\n* The full audio encoder (24 layers, d_model 1024, 16 heads, 4096 ff)\n* The Qwen3 1.7B LLM (28 layers, d_model 2048, 16 heads / 8 KV heads, 6144 ff, 152K vocab, RoPE θ=1e6)\n* The full GPT-2-style BPE vocab + merges, the audio mel filterbank, and the Hann window — everything the C++ side needs to run inference end-to-end without re-fetching the original safetensors.\n\n## Use with CrispASR (no Python at runtime)\n\n```bash\n# Build crispasr (one-time, no Python deps for runtime — only the converter\n# was Python).\ngit clone https://github.com/CrispStrobe/CrispASR\ncd CrispASR\ncmake -B build -DCMAKE_BUILD_TYPE=Release\ncmake --build build -j$(nproc) --target whisper-cli\n\n# Auto-download the recommended quant on first use:\n./build/bin/crispasr --backend qwen3 -m auto -f my_audio.wav\n\n# Or point at a local file:\n./build/bin/crispasr --backend qwen3 \\\n    -m qwen3-asr-1.7b-q4_k.gguf \\\n    -f my_audio.wav\n\n# Auto language detection (whisper-tiny LID pre-step):\n./build/bin/crispasr --backend qwen3 \\\n    -m qwen3-asr-1.7b-q4_k.gguf \\\n    -f my_audio.wav -l auto\n\n# Speech translation to German via the system-prompt instruction path:\n./build/bin/crispasr --backend qwen3 \\\n    -m qwen3-asr-1.7b-q4_k.gguf \\\n    -f my_audio.wav --translate -tl de\n\n# Word-level timestamps via the canary-ctc-aligner second pass:\n./build/bin/crispasr --backend qwen3 \\\n    -m qwen3-asr-1.7b-q4_k.gguf \\\n    -f my_audio.wav -am canary-ctc-aligner-q5_0.gguf -osrt -ml 1\n```\n\nPure C++ runtime: no Python, no PyTorch, no ONNX Runtime. The only build dep beyond a C++17 compiler and CMake is `libcurl` or `wget` for the auto-download path.\n\n## Languages\n\nSame as upstream Qwen3-ASR: 30 languages plus 22 Chinese dialects, including English, Chinese, Cantonese, German, French, Spanish, Japanese, Korean, Russian, Arabic, Hindi, and most major European languages. See the [base model card](https://huggingface.co/Qwen/Qwen3-ASR-1.7B) for the full matrix and accuracy figures.\n\n## How it was made\n\n```bash\n# 1. Download the base model from HF\nhf download Qwen/Qwen3-ASR-1.7B --local-dir ./Qwen3-ASR-1.7B\n\n# 2. Convert to F16 GGUF\npython models/convert-qwen3-asr-to-gguf.py \\\n    --input ./Qwen3-ASR-1.7B \\\n    --output qwen3-asr-1.7b-f16.gguf\n\n# 3. Quantize\n./build/bin/crispasr-quantize qwen3-asr-1.7b-f16.gguf qwen3-asr-1.7b-q8_0.gguf q8_0\n./build/bin/crispasr-quantize qwen3-asr-1.7b-f16.gguf qwen3-asr-1.7b-q4_k.gguf q4_k\n```\n\nThe converter is shared with the 0.6B variant — it reads all sizes (audio encoder layers, LLM hidden size, KV head count, vocab size, head dim) from `config.json` rather than hardcoding them, so the same script handles both checkpoints. Total: 708 tensors (348 F16, 360 F32) per file.\n\n## Bit-identical regression check\n\nBoth quants converge on samples/jfk.wav under CrispASR's qwen3 backend:\n\n```\nqwen3-asr-1.7b-f16  → \"And so, my fellow Americans, ask not what your country can do for you. Ask what you can do for your country.\"\nqwen3-asr-1.7b-q8_0 → \"And so, my fellow Americans, ask not what your country can do for you. Ask what you can do for your country.\"\nqwen3-asr-1.7b-q4_k → \"And so, my fellow Americans, ask not what your country can do for you; ask what you can do for your country.\"\n```\n\n## License\n\nApache-2.0, same as the upstream Qwen3-ASR-1.7B model.\n\n## Citation\n\n```bibtex\n@misc{qwen3asr,\n    title  = {Qwen3-ASR},\n    author = {Qwen Team},\n    year   = {2026},\n    url    = {https://huggingface.co/Qwen/Qwen3-ASR-1.7B}\n}\n```\n",
    "related_quantizations": []
  },
  "tags": [
    "ggml",
    "gguf",
    "asr",
    "speech-to-text",
    "audio",
    "transcription",
    "speech-llm",
    "qwen3",
    "automatic-speech-recognition",
    "zh",
    "en",
    "yue",
    "ar",
    "de",
    "fr",
    "es",
    "pt",
    "id",
    "it",
    "ko",
    "ru",
    "th",
    "vi",
    "ja",
    "tr",
    "hi",
    "ms",
    "nl",
    "sv",
    "da",
    "fi",
    "pl",
    "cs",
    "fil",
    "fa",
    "el",
    "ro",
    "hu",
    "mk",
    "base_model:Qwen/Qwen3-ASR-1.7B",
    "base_model:quantized:Qwen/Qwen3-ASR-1.7B",
    "license:apache-2.0",
    "region:us"
  ],
  "likes": 1,
  "downloads": 387,
  "gated": false,
  "private": false,
  "last_modified": "2026-04-11T17:41:11.000Z",
  "created_at": "2026-04-11T17:37:52.000Z",
  "pipeline_tag": "automatic-speech-recognition",
  "library_name": "ggml"
}

{
  "_id": "69da86f0e7d734b4eec72601",
  "id": "cstr/qwen3-asr-1.7b-GGUF",
  "modelId": "cstr/qwen3-asr-1.7b-GGUF",
  "sha": "4593c9d871d43870c73b3500723399409f31a3de",
  "createdAt": "2026-04-11T17:37:52.000Z",
  "lastModified": "2026-04-11T17:41:11.000Z",
  "author": "cstr",
  "downloads": 387,
  "likes": 1,
  "gated": false,
  "private": false,
  "pipeline_tag": "automatic-speech-recognition",
  "library_name": "ggml",
  "siblings_count": 5
}