SourcesQwen Team
publisher · 1 source
Qwen Team
1 source read from this publisher, newest first, with the methods each one describes.
Qwen3.5-Omni Technical Report
technical report · 2026-04-17 · 28 methods
Gated DeltaNetOn-Policy DistillationChunked prefillContinued pretrainingSpecialist Distillation128-channel mel-spectrogram audio front endAdaptive Rate Interleave Alignment (ARIA)Audio Transformer (AuT)Byte-level byte-pair encodingCausal streaming ConvNet codec decodercross-lingual voice cloningcustom-voice speech generationDirect Preference Optimization (DPO)Dynamic Attention Window Size TrainingExplicit text-string timestampsfour-stage Talker training pipelineGSPOHybrid Mixture of ExpertsInteraction-Aligned RLLightweight speaker fine-tuningMulti-Token Prediction module for residual codebooksmultilingual speech generationRVQ-based speech representationTalker system prompt for voice characteristicsTemporal-Modality Rotary Position EmbeddingThinker–Talker architectureThree-stage pre-training curriculumThree-stage Thinker post-training strategy