Confucius4-R2T2 is a streaming automatic speech recognition model from NetEase Youdao, built on Alibaba's Qwen3-ASR (base model Qwen/Qwen3-ASR-1.7B) and released as a sibling to the org's Confucius4-TTS model. Unlike systems that revise a live transcript as more audio arrives, R2T2 uses an append-only output mode: once text is committed, it is never rewritten. It is trained on stable-prefix data, forced time-alignment data, and token-level audio segmentation, with a Longest Stable Prefix (LSP) learning paradigm deciding when a prefix can safely be emitted; the tech report on this method has not been released yet. Decoding chunk size is configurable from 80 ms to 2 s, and the authors report 200-600 ms average latency with accuracy close to offline recognition. In the authors' benchmark tables at a 160 ms chunk setting, R2T2 improves sharply on Qwen3-ASR's own true-streaming baseline at the same setting (11.37% vs 24.79% WER on AMI, 2.13% vs 22.30% on LibriSpeech-clean, 5.87% vs 19.79% CER on Wenet-net) and posts lower error than the open-source X-ASR, WhisperRT, Nemotron and Voxtral baselines on every listed test set. It trails Qwen3-ASR's pseudo-streaming 2 s configuration, whose partial output can still be revised, on all sets, and AssemblyAI and two unnamed commercial services on several. The model targets Chinese and English, with unbenchmarked cross-lingual streaming on other languages including French, German, Japanese and Korean. The repository ships a vLLM backend for high-throughput inference (a Transformers backend is also available) and a WebSocket server for multi-client streaming that uses a separately downloaded FireRedVAD Stream-VAD model. The recommended setup runs on the official Qwen3-ASR Docker image with the NVIDIA Container Toolkit; Conda and uv source installs need Python 3.10+. A single BF16 checkpoint of about 2.04B parameters is published. Code is Apache-2.0; the weights carry NetEase Youdao's own Model Use License Agreement, which requires a separate license above 100 million monthly active users or RMB 1 billion in annual revenue, bars using the model to improve other AI models except non-commercial ones, and excludes high-risk deployments such as medical diagnosis and autonomous driving.