Confucius4-R2T2
NetEase Youdao's streaming speech recognizer that never revises words it has committed, adapted from Qwen3-ASR-1.7B.
In short. Confucius4-R2T2 turns speech into text as it is spoken, and the words it shows never change afterwards. That avoids the flicker you see in many live captions, and it lets software act on a transcript at once. It works best for Chinese and English, and it is free to use commercially below some large company limits.
Summary
Confucius4-R2T2 is a streaming speech recognizer from NetEase Youdao. R2T2 stands for Real Real-Time Transcription, and the model commits transcript text permanently, without going back to revise words it has already shown. It adapts Qwen3-ASR-1.7B with what NetEase calls Longest Stable Prefix learning on 1,500 hours of Chinese and English audio. Two training stages, supervised fine-tuning and GRPO, update only the language decoder, while the audio encoder and projector stay frozen. Decoding chunks can be set from 80 ms to 2 s, and a provisional-token setting trades accuracy against lag, with U=1 favoring quality and U=0 favoring lower latency.
The numbers are NetEase's own. Across 13 English and Chinese test sets, at 160 ms chunks with one provisional token, the average error rate is 5.76%, against 21.01% for the unadapted base model run the same way and 5.15% for a full-context reference. On LibriSpeech clean it scores 1.91% word error rate against 3.71% for Nemotron 3.5 and 3.05% for Voxtral RT, as the card labels them, and the card says it ranks first among the open-source models it tested. The same table has two anonymous commercial systems beating it on several sets, especially Chinese meeting speech. On 18 test sets in ten languages absent from post-training, its macro error rate is 8.40%, against 5.99% for the base model run offline.
The checkpoint is a 4.08 GB bf16 file. A separate GGUF repository adds Q4_K_M at 1.11 GB and Q8_0 at 1.83 GB plus audio projector files. It runs through vLLM on the official Qwen3-ASR Docker image, through a Python API, through a Transformers offline route or through llama.cpp decoding, and it takes topic and vocabulary hints. The repository also ships a WebSocket server for several clients at once, which uses a separate FireRedVAD voice-activity model. The code is Apache 2.0, and the weights use the NetEase Youdao Model Use License Agreement. The Hugging Face repository was created on 10 September 2026 and the paper, arXiv 2611.06043, is from November, so we give the month only. In the 30 days to 11 November 2026, Hugging Face counted about 45,000 downloads of the main repository and 6,500 of the GGUF one.
Good for
- Live captions and voice interfaces where words must not change after they appear
- Voice agents that act on a transcript the moment it arrives
- Chinese and English transcription on your own servers, built on an open Qwen3-ASR base
- Tuning the balance between accuracy and delay with the chunk size and provisional-token setting
Watch out for
- The accuracy tables are NetEase's own, and two commercial systems in them did better on several test sets
- Post-training used only Chinese and English, and the other ten languages tested trail the base model run offline
- The license needs a separate agreement above 100 million monthly users or RMB 1 billion of annual revenue, and it bars using the model to improve other commercial AI models
- The documented setup is built around vLLM or a bundled llama.cpp decoder with CUDA, and the card recommends Docker
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Memory figures for each variant of Confucius4-R2T2. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.
| Variants | Parameters | Download size | Memory needed | Fits on |
|---|---|---|---|---|
| Confucius4-R2T2 BF16 | 2.04B | 4.08 GB | 4.69 GB | Laptop or phone |
Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.
The license in plain words
The license for NetEase Youdao's Confucius4-R2T2 speech recognizer. Commercial use is licensed unless you or an affiliate had more than 100 million monthly active users or more than RMB 1 billion of annual revenue, and then you need a separate license. It also bars using the model to improve other AI models, apart from non-commercial ones, and prohibits high-risk uses.
- Free commercial use only below a revenue cap
- Acceptable-use policy applies
More about NetEase Youdao Model Use
See also
- Best open-source speech models for transcription and text-to-speech
- Best open-source speech recognition (speech-to-text) models in 2026
Frequently asked questions
- What is Confucius4-R2T2?
- Confucius4-R2T2 turns speech into text as it is spoken, and the words it shows never change afterwards. That avoids the flicker you see in many live captions, and it lets software act on a transcript at once. It works best for Chinese and English, and it is free to use commercially below some large company limits.
- Can I use Confucius4-R2T2 commercially?
- Yes, with conditions. Confucius4-R2T2 is released under the NetEase Youdao Model Use, which allows commercial use but adds extra rules: Free commercial use only below a revenue cap; Acceptable-use policy applies. Read the license text before you ship a product.
- Is Confucius4-R2T2 open source?
- Confucius4-R2T2 is open-weight, not open source in the strict sense. You can download the model files, but the NetEase Youdao Model Use adds conditions that the Open Source Initiative definition does not allow.
- Can I run Confucius4-R2T2 on my own computer?
- Yes. The smallest variant, Confucius4-R2T2, needs about 4.69 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was Confucius4-R2T2 released?
- Confucius4-R2T2 was released by NetEase Youdao on September 2026.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Confucius4-R2T2 model card (Hugging Face) huggingface.co
- Confucius4-R2T2 repository: install, backends and evaluation (GitHub) github.com
- NetEase Youdao Model Use License Agreement (GitHub) github.com
- Confucius4-R2T2 GGUF repository (Hugging Face) huggingface.co
- Confucius4-R2T2 paper (arXiv 2611.06043) arxiv.org
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Qwen3-ASR Alibaba Qwen | 918M – 2.35B | – | Apache 2.0 (Permissive) | 2.12 GB Laptop or phone | |
| Voxtral Realtime Mistral AI | 4.43B | 128K | Apache 2.0 (Permissive) | 10 GB Laptop or phone | |
| Nemotron 3.5 ASR NVIDIA | 638M | – | OpenMDW-1.1 (Permissive) | 2.93 GB Laptop or phone | |
| Nemotron 3.5 Transcribe NVIDIA | 34B 4.3B active | 64K | OpenMDW-1.1 (Permissive) | 25 GB Gaming GPU |