netease-youdao/Confucius4-T3PO
Confucius4-T3PO is a 14.8 billion parameter text-to-text simultaneous machine translation (SiMT) model developed by the NetEase Youdao AI team. It supports streaming text input and real-time READ/WRITE decisions, dynamically deciding whether to wait for more context or produce an incremental translation. This model excels at Chinese-to-English and English-to-Chinese simultaneous translation, offering adjustable latency modes for varying quality-latency requirements. It is designed for live text or speech translation where incremental output and bounded latency are critical.
Loading preview...
Confucius4-T3PO: Simultaneous Machine Translation
Confucius4-T3PO is a 14.8 billion parameter text-to-text simultaneous machine translation (SiMT) model from NetEase Youdao. It operates on a three-stage pipeline involving high-quality segment-aligned data construction, streaming cold-start, and Pareto-aware reinforcement learning for joint quality–latency optimization. The model processes fine-grained text chunks, making dynamic READ/WRITE decisions to either wait for more context or immediately produce an incremental translation. It also utilizes an interleaved history protocol for KV-cache reuse, reducing computational overhead.
Key Capabilities
- Fully Streaming Text Translation: Supports fine-grained character and word-level input, with committed translations being append-only and never rewritten. Maintains a stable prefix for KV-cache reuse.
- Adjustable Latency Modes: Offers flexible switching across multiple quality–latency tiers, adapting to different simultaneous interpretation scenarios from low-latency to high-quality.
- Retained General Instruction-Following Ability: Preserves the instruction-following capabilities of its Qwen base model, allowing for further extensions like terminology constraints.
- Cross-Lingual Generalization: Demonstrates some generalization, with observed support for Chinese-to-Japanese streaming translation, though primary focus is Chinese and English.
- Optimized Quality–Latency Frontier: Employs a frontier-aware reinforcement learning algorithm to advance the Pareto frontier for simultaneous translation, balancing quality and latency.
Use Cases
Confucius4-T3PO is intended for live Chinese–English text or speech translation where incremental output and bounded latency are crucial. While it is a text-to-text model, it can be cascaded with an external streaming ASR model like R2T2 for a complete speech-to-text (S2T) pipeline. The model's ability to dynamically adjust latency makes it suitable for various real-time interpretation scenarios.