Qwen3-ASR-1.7B-ViMD-Full — GGUF

Community GGUF conversions of duclvQ/qwen3-asr-1.7b-vimd-full, a Vietnamese fine-tuned Qwen3-ASR-1.7B model.

This repository distributes converted weights only; it does not introduce a new fine-tuned checkpoint. Two GGUF families are provided for different inference runtimes.

Models and downloads

Compatibility: Select the GGUF matching your runtime. The qwen3-asr.cpp and CrispASR exports are not interchangeable merely because both use .gguf.

All five files originate from the ViMD Full BF16 fine-tuned checkpoint. F16 is an FP16 conversion of BF16, not a bit-exact BF16 copy. Q8_0 and Q4_1 are GGUF block quantizations, not the HQQ checkpoints in the upstream repository.

Conversion and runtime compatibility

qwen3-asr.cpp exports

• Converted using predict-woo/qwen3-asr.cpp and its convert_hf_to_gguf.py.
• 707 tensors converted; 0 skipped. GGUF headers and tensor metadata were verified.
• Three audio-encoder convolution tensors remain F16 in the Q8_0 and Q4_1 versions because their shapes were not supported by the selected block quantization.
• Intended for qwen3-asr.cpp, Subtitle Edit with the Qwen3 ASR CPP integration, and compatible implementations. End-to-end compatibility and transcription quality have not been independently established for all runtimes.

CrispASR exports

• Converted using the CrispStrobe/CrispASR Qwen3-ASR conversion path; Q8_0 was produced with crispasr-quantize.
• Each file has 710 GGUF tensor entries, including audio-preprocessing arrays and an explicitly materialized LM output projection, output.weight, as expected by the CrispASR loader.
• F16: 4,704,800,576 bytes. Q8_0: 2,506,723,200 bytes. Other tensor types can remain F16/F32 where required.
• These are alternate exports of the same source checkpoint, not an additional fine-tune.

CrispASR Q8_0 smoke test: The model loaded and transcribed a 9.2-second English speech clip using CrispASR v0.8.49 on Windows x86_64, with Vulkan selected on an Intel Iris Xe (i5-1135G7) and 4 CPU threads. Reported processing time was 8.59 seconds (~1.07× real time) for this clip. This confirms the observed load-and-transcribe path; it is not a GPU/CPU benchmark or Vietnamese accuracy evaluation.

Example (CrispASR):

crispasr --backend qwen3 --gpu-backend vulkan -dev 0 -m "qwen3-asr-1.7b-vimd-crisp-q8_0.gguf" -f "audio.wav" -l vi -t 4 -v

Four-way GGUF tensor analysis

Metadata inspection compared the original Qwen3-ASR-1.7B Q8_0 export with three ViMD Full qwen3-asr.cpp exports. It compares tensor names, shapes, types, counts, and encoded sizes—not the actual weight values or recognition quality.

Why is ViMD Q8_0 about 593.50 MiB smaller?

The original Q8_0 export is 622,330,336 bytes (593.50 MiB) larger. The difference is almost entirely one F16 output.weight tensor:

All 707 shared tensor names have matching shapes, types, and encoded byte sizes. The extra tensor contributes 622,329,856 bytes; the remaining 480 bytes are non-tensor overhead/alignment.

output.weight is the language-model output projection (LM head). Its omission is consistent with weight tying (reusing input token embeddings), but metadata inspection does not prove that weights are numerically equivalent or that both runtimes interpret them identically. A smaller file does not establish faster inference, less peak RAM use, or better ASR accuracy.

Encoded storage by tensor type

GiB/MiB use binary units. These are on-disk sizes, not runtime memory measurements. Analysis of four local files was recorded on 2026-11-01 at 19:02:32 UTC.

Validation status and limitations

SHA256 checksums

qwen3-asr.cpp exports

e65273ea884c4fd55b7662ddf3aa23d5f80bf736b0455926cfd2ade241d88ba0  qwen3-asr-1.7b-vimd-f16.gguf
0ee16ae94d32712d7d3b41a9bd3ff533a0c8b4ec24b71a7b61ae95b3a5b24fc3  qwen3-asr-1.7b-vimd-q8_0.gguf
6235c03492ff9577c3baa02160957817edea15b879430cc6b4986047fcc87ce0  qwen3-asr-1.7b-vimd-q4_1.gguf

CrispASR exports (local files used for upload)

9596875a792fe05f5448a0825a7813e872d3f03ac67fac6ac994e8c5ca92403d  qwen3-asr-1.7b-vimd-crisp-f16.gguf
f7aaf09d5241b8191cb2326f90f0b0511e19c6fa56fb590bf9b8f5fe0ced6115  qwen3-asr-1.7b-vimd-crisp-q8_0.gguf

Credits and license

• Base architecture: Qwen team.
• Original ViMD Full fine-tuning: duclvQ.
• Conversion tools: predict-woo/qwen3-asr.cpp and CrispStrobe/CrispASR.
• GGUF conversion, quantization, packaging, and tensor analysis: BaoKhuong.

The upstream repository declares Apache-2.0. Review the source model and ViMD dataset licenses before redistribution or commercial deployment.