netease-youdao/Confucius4-R2T2

NetEase Youdao releases Confucius4-R2T2, a streaming speech recognition model built on Qwen3-ASR. It supports 80 ms to 2 s decoding chunks with append-only output, ensuring committed text remains unchanged. The model provides vLLM and Hugging Face backends for inference, targeting real-time captioning and agent pipelines with low latency.