CoolFace
Modelpublic

bearcove/qwen3-asr-1.7b-tq6_1s-text-tq4_1s-audio

sourceHugging Faceapache-2.0updated 5mo agoView on Hugging Face
0likes6downloads
Model Card

Qwen3-ASR 1.7B TQ61S Text / TQ41S Audio

This repository contains a TurboQuant quantization of `Qwen/Qwen3-ASR-1.7B`.

The model card and license follow the upstream Qwen3-ASR model. The weight file uses a packed TurboQuant safetensors layout and requires a runtime that understands that layout; it is not a drop-in replacement for the upstream transformers checkpoint.

Quantization

  • —Base model: Qwen/Qwen3-ASR-1.7B
  • —Text linear weights: TQ6_1S
  • —Audio tower linear weights: TQ4_1S
  • —Non-linear parameters, biases, normalization weights, tokenizer, and config: retained from the upstream checkpoint where applicable
  • —Text hidden size: 2048
  • —Text layer count: 28
  • —Hugging Face repo: bearcove/qwen3-asr-1.7b-tq6_1s-text-tq4_1s-audio

TurboQuant weight format

This checkpoint uses TurboQuant for model weights. The packed tensors are not ordinary int4/int8 matrices, and the 1S suffix is part of the stored basis.

Quantized linear weights are divided into 32-value blocks. For each block:

  1. 1.The 32 source values are transformed with a fixed signed 32-lane Walsh-Hadamard rotation.
  2. 2.The rotated block is split into two 16-value halves.
  3. 3.Each half gets its own fp16 scale.
  4. 4.Each rotated value is represented by the nearest entry in a fixed, non-uniform centroid table.
  5. 5.The file stores the two fp16 scales plus packed centroid indices.

For TQ6_1S, each 32-weight block is stored as 4 bytes of scales plus 24 bytes of 6-bit centroid indices, for 28 bytes total. For TQ4_1S, each 32-weight block is stored as 4 bytes of scales plus 16 bytes of 4-bit centroid indices, for 20 bytes total.

Dequantization reconstructs the rotated values as scale * centroid, then applies the inverse signed rotation to return to the normal weight basis. A runtime that reads these weights must implement that inverse transform; treating the packed bytes as plain affine int4/int6 weights will produce incorrect results.

Safetensors layout

model.safetensors includes both packed TurboQuant tensors and passthrough BF16 tensors. The quantized payloads are stored as U8 tensors, and the safetensors metadata contains a turboquant.index entry describing the quantized weights and their formats.

Text linear weights use split tensors:

  • —<weight>.scales: packed fp16 scale bytes
  • —<weight>.qs: packed centroid-index bytes

Audio linear weights use packed block tensors:

  • —<weight>.blocks: interleaved scale and centroid-index bytes

Parameters that are not quantized, such as normalization weights and other non-linear or structural tensors, remain in their upstream dtype where applicable.

License

This quantized checkpoint is distributed under the same license as the upstream Qwen3-ASR model: Apache License 2.0.

References

Links

Acknowledgements

This checkpoint is a quantized derivative of Qwen3-ASR. All credit for the upstream model architecture, training, tokenizer, processor configuration, and model card belongs to the Qwen team.

bearcove/qwen3-asr-1.7b-tq6_1s-text-tq4_1s-audio · CoolFace