PublishersQwen Team

Qwen Team

1 document read from this publisher.

Qwen3.5-Omni Technical Report

technical report · 2026-04-17 · 28 techniques

Gated DeltaNet On-Policy Distillation Chunked prefill Continual pretraining Specialist Distillation Adaptive Rate Interleave Alignment (ARIA)Audio Transformer (AuT)Byte-level Byte Pair Encoding (BPE)Causal streaming ConvNet codec decoder cross-lingual voice cloning custom-voice speech generation Direct Preference Optimization Dynamic Attention Window Size Training Explicit text-string timestamps four-stage Talker training pipeline GSPO Hybrid Mixture of Experts Interaction-Aligned Reinforcement Learning Lightweight speaker fine-tuning Mel-spectrogram audio front end Multi-Token Prediction for residual codebooks multilingual speech generation Residual Vector Quantization (RVQ) speech representation Talker system prompt for voice characteristics Temporal-Modality Rotary Position Embedding Thinker–Talker architecture Three-stage pre-training curriculum Three-stage Thinker post-training strategy