Qwen3-Omni Captioner

Overview

Qwen3-Omni Captioner is a state-of-the-art, natively end-to-end multilingual omni-modal foundation model developed by Alibaba Cloud’s Qwen team. It supports diverse input modalities including text, images, audio, and video, and delivers real-time streaming responses in both natural text and speech. Qwen3-Omni maintains high performance across modalities without any degradation, making it a leading multimodal AI solution.

Technical Specifications

• Thinker-Talker Architecture: Qwen3-Omni features a Thinker-Talker design, where the Thinker generates text and the Talker converts the Thinker's high-level representations into streaming speech tokens. This separation enables specialized processing for text generation and real-time audio synthesis.
• Ultra-Low-Latency Streaming: To achieve real-time streaming voice output, the Talker predicts multi-codebook sequences autoregressively. At each step, a Multi-Token Predictor (MTP) module outputs residual codebooks for the current audio frame. The Code2Wav renderer then incrementally synthesizes the waveform, producing audio frame-by-frame for seamless streaming.
• AuT Audio Encoder: The model’s audio encoder, AuT, is trained on 20 million hours of audio data, providing strong and generalizable audio feature extraction.
• MoE Architecture: Both Thinker and Talker subsystems use Mixture-of-Experts (MoE) models for high concurrency and fast inference, activating only a subset of parameters per token for efficiency.

Performance Benchmarks

Qwen3-Omni achieves state-of-the-art results on 22 out of 36 audio and audio-visual benchmarks, outperforming strong closed-source models including Gemini 2.5 Pro and GPT-4o-Transcribe.

• Text Understanding: Competitive with top models on MMLU, GPQA, reasoning, and code tasks.
• Audio Recognition (ASR): Word Error Rate (WER) on par or better than Seed-ASR and GPT-4o-Transcribe across numerous datasets.
• Multimodal Reasoning: Strong performance on audio-visual question answering and video description benchmarks.
• Speech Generation: High-quality multilingual speech synthesis with consistent speaker identity across 10 languages.
• Streaming Latency: Low first-packet latency (~211 ms), enabling near-instant speech response.
• Audio Captioning: Fine-tuned model excels in generating detailed, accurate captions for arbitrary audio.

Key Features

• Architecture: MoE-based Thinker–Talker design utilizing Audio Transformer (AuT) pretraining and multi-codebook speech synthesis for low-latency and high-fidelity output.
• Extensive Reasoning: Enhanced reasoning ability across modalities with the Thinking model variant.
• Customization: Behavior customizable via system prompts to control tone and style of interaction.
• Open-Source Audio Captioner: Fine-tuned Qwen3-Omni-30B-A3B-Captioner provides detailed, low-hallucination audio descriptions.
• Real-Time Interaction: Supports natural turn-taking with immediate text or speech responses.

API Pricing

• Input $4.953
• Output $3.978

Use Cases

• Multilingual chatbots with audio and visual understanding
• Real-time streaming transcription and translation across languages
• Detailed audio and video content analysis and captioning
• Multimodal question answering and reasoning
• Voice assistants with natural speech and multimodal comprehension
• Interactive multimedia content generation and navigation

Code Sample

Comparison with Other Models

vs Gemini 2.5 Pro: Qwen3-Omni matches or exceeds Gemini’s performance on audio-video benchmarks and offers better open-source accessibility. Comparable ASR performance with lower latency in streaming speech generation.

vs Seed-ASR: Qwen3-Omni achieves superior or comparable word error rates while providing broader multimodal capabilities beyond audio.

vs GPT-4o: Qwen3-Omni excels in multimodal audio and video tasks while maintaining strong text tasks proficiency. Lower latency streaming audio output with native multi-codebook speech codec.

API Integration

Accessible via AI/ML API. Documentation: available here.