File size: 3,521 Bytes
edcd0cc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
---
license: other
license_name: netease-youdao-model-use-license-agreement
license_link: https://raw.githubusercontent.com/netease-youdao/Confucius4-R2T2/refs/heads/master/MODEL_LICENSE
base_model:
  - netease-youdao/Confucius4-R2T2
pipeline_tag: automatic-speech-recognition
language:
  - zh
  - en
tags:
  - FastFlowLM
  - NPU
  - AMD
  - Ryzen-AI
  - asr
  - speech-recognition
  - qwen3-asr
---

# Confucius4-R2T2 for AMD Ryzen™ AI NPUs (FastFlowLM)

[Confucius4-R2T2](https://huggingface.co/netease-youdao/Confucius4-R2T2) by NetEase Youdao ("Real Real-Time Transcription", fine-tuned from
[Qwen3-ASR-1.7B](https://huggingface.co/Qwen/Qwen3-ASR-1.7B)) converted for [FastFlowLM](https://github.com/FastFlowLM/FastFlowLM) (FLM) on the XDNA 2 NPU.
2B parameters, 30 languages (best at Chinese and English).

> Any modifications made to the original model in this Derivative Work are not endorsed, warranted, or guaranteed by the original right-holder of the original model, and the original right-holder disclaims all liability related to this Derivative Work.

This is a **converted copy** of the original weights, subject to the **NetEase Youdao Model Use License Agreement** (`MODEL_LICENSE`, included). Please read it: among other
terms it requires a separate license from NetEase Youdao for organizations with more than 100 million monthly active users or more than RMB 1 billion annual revenue
(including their affiliates), and it forbids use of the model or its outputs to improve other AI models except non-commercial ones.

## Run

Requires FastFlowLM 1.0.14 or newer.

```shell
flm run r2t2:1.7b
# then:  /input "audio.wav"            (wav, flac, mp3, ogg)
#        /input "audio.wav" hot words, names, topic        (text after the file is context for the model)

flm serve r2t2:1.7b
curl http://localhost:52625/v1/audio/transcriptions -F file=@audio.wav -F model=r2t2:1.7b
```

Whole-file transcription, or streaming with `-F stream=true` (append-only text while the audio is processed, `chunk_size_ms` 80-2000).
See the [model page](https://fastflowlm.com/docs/models/r2t2/) and the [benchmarks](https://fastflowlm.com/docs/benchmarks/r2t2_results/).

## What is in this package

| File | Content |
|---|---|
| `model.q4nx` | text decoder (Qwen3-1.7B shape): **Q8_0** projections and output head, BF16 embedding (the CPU prompt prefill reads it) |
| `npu_q41.q4nx` | **Q4_1** copies of q/k/o/gate/up and the output head for the NPU decode (v and down stay Q8_0 from `model.q4nx`) |
| `audio_weight.q4nx` | audio encoder (24 layers), BF16, in the NPU matmul tile layout |
| `config.json`, `tokenizer.json`, `tokenizer_config.json`, `chat_template.jinja` | from the original checkpoint |
| `MODEL_LICENSE` | NetEase Youdao Model Use License Agreement (original) |

The NPU decode reads mixed precision: 4-bit everywhere (Q4_0, Q4_1, Q4_K round-to-nearest) breaks this fine-tune (empty
transcripts or Mandarin answered in English), but Q4_1 for q/k/o/gate/up and the output head with v_proj and down_proj at Q8_0
keeps its output - and decodes ~24 % faster than Q8_0 throughout. On 120 English (LibriSpeech test-clean) and 60 Mandarin
(FLEURS) utterances, compared with the original BF16 model, FLM gives WER 2.81 % vs 2.89 % (English) and CER 6.37 % vs 8.96 %
(Mandarin; the gap is mostly how numbers are written, digits vs Chinese numerals, which the scoring does not normalise).

## Source

Converted from `netease-youdao/Confucius4-R2T2` (BF16, revision `185ce639`) with the FastFlowLM q4nx converter.