Alibaba’s Qwen team has made Qwen3-Omni fully open source, enabling free commercial use of a multimodal AI model that rivals proprietary systems from OpenAI and Google.
Alibaba’s Qwen team has unveiled Qwen3-Omni, a natively end-to-end omni-modal AI model capable of processing text, image, audio, and video inputs, with outputs in text and audio. Available under an Apache 2.0 open source license, Qwen3-Omni allows developers and enterprises to freely download, modify, and deploy the model commercially, setting a new benchmark in open access multimodal AI.
The model supports 119 languages for text, 19 for speech input, and 10 for speech output, including dialects such as Cantonese. Its Thinking Mode allows context lengths of up to 65,536 tokens with reasoning chains of 32,768 tokens. Distinct variants include the Instruct Model (full text, audio, video capabilities), Thinking Model (text-only reasoning), and Captioner Model (audio captioning with low hallucination).
Qwen3-Omni uses a Thinker–Talker architecture—the Thinker handles reasoning and multimodal understanding, while the Talker generates natural speech from audio-visual features. Its Mixture-of-Experts (MoE) design ensures high concurrency and fast inference, with streaming latency as low as 234 ms for audio and 547 ms for video. The model was pretrained on ~2 trillion tokens across text, audio, images, and video, with the Audio Transformer trained on 20M hours of audio.
Benchmark results show strong performance across text, reasoning, speech, and vision tasks, surpassing GPT-4o and Gemini 2.0 Flash in multiple metrics. Applications include multilingual transcription, translation, audio captioning, OCR, music tagging, video understanding, and real-time AI assistants for tech support. Enterprises can also customise behaviour via system prompts for persona, style, or domain-specific use.
Lin from Alibaba AI Research commented: “This might bring some changes to the landscape of open source Omni models! Hope you enjoy it!” Qwen3-Omni marks a major open source challenge to proprietary multimodal AI offerings from US tech giants.
