Confucius4-R2T2-gguf

Confucius4-R2T2 — GGUF (audio.cpp)

GGUF conversions of NetEase Youdao Confucius4-R2T2, a low-latency, append-only streaming ASR model (a Qwen3-ASR-1.7B fine-tune with Longest Stable Prefix decoding), ready to run with audio.cpp — no Python, no PyTorch, local inference on Metal / CUDA / CPU.

• Original model: netease-youdao/Confucius4-R2T2 · source code · demo
• Runtime: audio.cpp (ggml), family confucius4_r2t2

Runtime compatibility

These files use the confucius4_r2t2 family and embed its schema-v1 model specification. Use a build containing the family rename and streaming-delta fix from audio.cpp PR #604 (validated at commit 40dec7147128649a4180a5c982a7808cd272fc3e). Older builds that only register r2t2_asr need to be updated. Model-manager package IDs and session-option prefixes use confucius4_r2t2 as shown below; the GGUF filenames remain unchanged.

Both precisions were regenerated from the original upstream safetensors using the updated converter and explicit model spec. Tensor names, shapes, types, and weight payloads are unchanged from the previous GGUF files; the update refreshes embedded package metadata. This is not a model retraining. VAD endpointing is separate work in progress and is not enabled by this update.

Files

Both files are self-contained: the tokenizer, processor config, generation config, chat template and the audio.cpp model spec are embedded, so the runtime needs nothing but the binary and the single .gguf (the embedded files are unpacked into a temporary directory at load time). The embedded spec carries the schema-v1 option contract, so option validation comes from the file itself rather than from a spec installed next to the runtime.

Precision policy: Q8_0 and higher are supported. 4-bit and 5-bit quantizations (legacy and k-quant) are rejected at load time with an actionable error, because this graph's kernels are not validated below Q8_0 and otherwise decode to empty text.

SHA-256:

19f5ccd624484bcb5d44301437de41560b0ecc40c430e8850dfeefefbe82ccf5  r2t2-q8_0.gguf
d1b531ceaf5640d98352d3a9180238d99d36d393e160afd4692031077e7bae2c  r2t2-f16.gguf

Usage

Install

# via the audio.cpp model manager
python3 tools/model_manager_v2.py install confucius4_r2t2_q8_0     # or confucius4_r2t2_f16

# or directly
hf download davidxifeng/Confucius4-R2T2-gguf r2t2-q8_0.gguf \
  --local-dir models/Confucius4-R2T2-GGUF

Offline transcription

audiocpp_cli --task asr --family confucius4_r2t2 \
  --model models/Confucius4-R2T2-GGUF/r2t2-q8_0.gguf \
  --backend metal --audio speech_16k.wav --text-out transcript.txt

Streaming (real-time, append-only)

audiocpp_cli --task asr --mode streaming --family confucius4_r2t2 \
  --model models/Confucius4-R2T2-GGUF/r2t2-q8_0.gguf \
  --backend metal --audio speech_16k.wav \
  --session-option confucius4_r2t2.chunk_size_ms=320 \
  --text-out transcript.txt

Chunk sizes from 80 ms to 2 s are supported; 320 ms is a good default on Apple Silicon. --audio - streams raw 16 kHz mono PCM from stdin for live sources. Pass --language Chinese to skip language detection and --text "hotword, term" for a context/hotword prompt.

Server

{
  "id": "r2t2-asr",
  "family": "confucius4_r2t2",
  "path": "models/Confucius4-R2T2-GGUF/r2t2-q8_0.gguf",
  "task": "asr",
  "mode": "streaming",
  "session_options": { "confucius4_r2t2.chunk_size_ms": "320" }
}

curl http://127.0.0.1:8488/v1/audio/transcriptions -F model=r2t2-asr -F [email protected]
curl -N http://127.0.0.1:8488/v1/audio/transcriptions \
  -F model=r2t2-asr -F stream=true -F [email protected]     # SSE transcript deltas

See the audio.cpp model documentation for the full option reference, the LSP streaming contract, and GGUF details.

Conversion recipe

audiocpp_gguf \
  --input /path/to/Confucius4-R2T2/model.safetensors \
  --root  /path/to/Confucius4-R2T2 \
  --family confucius4_r2t2 \
  --model-spec model_specs/confucius4_r2t2.json \
  --output r2t2-q8_0.gguf \
  --type q8_0

No weight-name or head flags are needed: the checkpoint ties the LM head to the token embedding (lm_head.weight is absent) and the loader detects that. Quantization only; no other weight modification.

Verification

The refreshed Q8_0 and F16 files are validated with the confucius4_r2t2 runtime against the audio.cpp English sample (assets/resources/sample_16k.wav): offline and final streaming transcripts must match the MPS reference golden, and the concatenated streaming deltas must be a nonempty prefix of the expected transcript. Automatic language detection is explicitly exercised.

Streaming deltas contain transcript text only. Rollback fragments such as language are suppressed and do not advance the published text offset. The uncommitted tail is returned in the final transcript, not as a final delta. The old reference behavior that leaked language metadata is not preserved.

Language support

Optimized for Chinese and English, with support for Cantonese, Arabic, German, French, Spanish, Portuguese, Indonesian, Italian, Korean, Russian, Thai, Vietnamese, Japanese, Turkish, Hindi, Malay, Dutch, Swedish, Danish, Finnish, Polish, Czech, Filipino, Persian, Greek, Romanian, Hungarian and Macedonian.

License and attribution

This is a quantized Derivative Work of NetEase Youdao Confucius4-R2T2, distributed under the NetEase Youdao Model Use License Agreement — full text in LICENSE (English) and LICENSE_zh (Chinese); see also NOTICE. The source code of the original model is Apache-2.0, the model weights are not.

Required statement (license section 4.1(a)):

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.

Key restrictions: a separate commercial license is required above 100 M monthly active users or RMB 1 B annual revenue (section 2.2); the model may not be used to improve other AI models except Confucius4-R2T2, its derivatives, or non-commercial models (section 3.4(c)); high-risk uses such as medical diagnosis, autonomous driving, military, critical-infrastructure control, large-scale biometric surveillance and automated decision-making are prohibited (section 4.2). No trademark rights are granted (section 5.2).

Citation

@misc{Confucius4-R2T2,
  title        = {Confucius4-R2T2: A Low Latency and High Accuracy Real-Time Speech Recognition Model},
  author       = {NetEase Youdao},
  year         = {2026},
  howpublished = {https://github.com/netease-youdao/Confucius4-R2T2}
}