Model · Alibaba · Realtime-Voice
Qwen3-Omni
Alibaba's open model for live voice conversation. It listens, sees images and video, and answers in text or natural speech in real time — an open alternative to the voice modes of ChatGPT and Gemini, but it needs data-centre hardware.
- Modality
- Realtime-Voice
- License
- Apache 2.0 (Open)
- Parameter size
- 30B total, 3B active (MoE)
- Released
- September 22, 2025
- Last verified
- October 28, 2026
- Runs locally
- Yes
- Also handles
- speech-to-text, text-to-speech, vision, text
What you need to run it
- Mac
- Yes, but slow. Any Apple Silicon Mac (M1 or newer) with at least 32 GB (64 GB recommended) of memory. It runs, but much slower than on an NVIDIA card.
- Windows / Linux PC
- NVIDIA only. A graphics card with at least 80 GB of video memory. AMD and Intel cards are not supported.
- No graphics card
- Not practical. Too slow without a graphics card or Apple Silicon.
Qwen's own voice-reply code needs about 80 GB of NVIDIA graphics memory; on a Mac you can run the text, image and audio-listening parts in MLX or llama.cpp, but spoken replies work only in the experimental MLX port.
Numbers are for: Official BF16 weights for the GPU numbers (voice in and out); Mac numbers are for the community 4-bit text/image/audio-in build (about 22 GB). Based on the maker’s own figures; where the maker gives only a graphics-card figure, the Mac memory is our estimate.
Strengths
- Apache 2.0 — no commercial restrictions
- One model for speech in, speech out, images and video
- 19 speech input languages and 10 speech output languages (119 for text)
- Low-latency streaming with natural turn-taking
- Thinking and Captioner variants published alongside the Instruct model
Weaknesses
- Needs a data-centre GPU: about 79 GB of GPU memory minimum in BF16 with Transformers
- Inference with Hugging Face Transformers is slow for this MoE model; vLLM is recommended
- No newer open Qwen Omni release since September 2025
Try it
| Where | Type | Notes |
|---|---|---|
| Hugging Face | weights | Apache 2.0 |
| GitHub | local | Transformers and vLLM recipes, Docker image |
| Alibaba Cloud Model Studio | hosted-api | Real-time API for Qwen3-Omni-Flash; API key required |
| Qwen Chat | product | Voice and video call option in the chat box |
Official sources
- Model card model
- GitHub github
Change log
- — Added hardware requirements (Mac support, memory, NVIDIA). Source: Qwen3-Omni GitHub README (79 GB BF16 table); mlx-community/Qwen3-Omni-30B-A3B-Instruct-4bit; ggml-org GGUF; llama.cpp issue 21956 (no audio output yet).
- — Initial entry. Added 2026-10-28 from the popularity sweep: 646k Hugging Face downloads/month for the Instruct model; 4.1k GitHub stars. Released 2025-09-22 per the GitHub README news. Licence is Apache 2.0 (GitHub LICENSE; the Hugging Face card lists it as 'other' with licence name apache-2.0).
Esc