--- license: other license_name: apache-2.0 language: - en tags: - multimodal library_name: transformers pipeline_tag: any-to-any base_model: - Qwen/Qwen3-Omni-30B-A3B-Thinking --- ## Performance | Metric | This model | Original model ([Qwen/Qwen3-Omni-30B-A3B-Thinking](https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Thinking)) | | :----- | :--------: | :---------------------------: | | **KL divergence** | 0.003563 | 0 *(by definition)* | | **Refusals** | 7/100 | 99/100 | # Qwen3-Omni

## Overview ### Introduction

Qwen3-Omni is the natively end-to-end multilingual omni-modal foundation models. It processes text, images, audio, and video, and delivers real-time streaming responses in both text and natural speech. We introduce several architectural upgrades to improve performance and efficiency. Key features: * **State-of-the-art across modalities**: Early text-first pretraining and mixed multimodal training provide native multimodal support. While achieving strong audio and audio-video results, unimodal text and image performance does not regress. Reaches SOTA on 22 of 36 audio/video benchmarks and open-source SOTA on 32 of 36; ASR, audio understanding, and voice conversation performance is comparable to Gemini 2.5 Pro. * **Multilingual**: Supports 119 text languages, 19 speech input languages, and 10 speech output languages. - **Speech Input**: English, Chinese, Korean, Japanese, German, Russian, Italian, French, Spanish, Portuguese, Malay, Dutch, Indonesian, Turkish, Vietnamese, Cantonese, Arabic, Urdu. - **Speech Output**: English, Chinese, French, German, Russian, Italian, Spanish, Portuguese, Japanese, Korean. * **Novel Architecture**: MoE-based Thinker–Talker design with AuT pretraining for strong general representations, plus a multi-codebook design that drives latency to a minimum. * **Real-time Audio/Video Interaction**: Low-latency streaming with natural turn-taking and immediate text or speech responses. * **Flexible Control**: Customize behavior via system prompts for fine-grained control and easy adaptation. * **Detailed Audio Captioner**: Qwen3-Omni-30B-A3B-Captioner is now open source: a general-purpose, highly detailed, low-hallucination audio captioning model that fills a critical gap in the open-source community. ### Model Architecture

### Cookbooks for Usage Cases Qwen3-Omni supports a wide range of multimodal application scenarios, covering various domain tasks involving audio, image, video, and audio-visual modalities. Below are several cookbooks demonstrating the usage cases of Qwen3-Omni and these cookbooks include our actual execution logs. You can first follow the [QuickStart](#quickstart) guide to download the model and install the necessary inference environment dependencies, then run and experiment locally—try modifying prompts or switching model types, and enjoy exploring the capabilities of Qwen3-Omni!

Category	Cookbook	Description
Audio	Speech Recognition	Speech recognition, supporting multiple languages and long audio.
	Speech Translation	Speech-to-Text / Speech-to-Speech translation.
	Music Analysis	Detailed analysis and appreciation of any music, including style, genre, rhythm, etc.
	Sound Analysis	Description and analysis of various sound effects and audio signals.
	Audio Caption	Audio captioning, detailed description of any audio input.
	Mixed Audio Analysis	Analysis of mixed audio content, such as speech, music, and environmental sounds.
Visual	OCR	OCR for complex images.
	Object Grounding	Target detection and grounding.
	Image Question	Answering arbitrary questions about any image.
	Image Math	Solving complex mathematical problems in images, highlighting the capabilities of the Thinking model.
	Video Description	Detailed description of video content.
	Video Navigation	Generating navigation commands from first-person motion videos.
	Video Scene Transition	Analysis of scene transitions in videos.
Audio-Visual	Audio Visual Question	Answering arbitrary questions in audio-visual scenarios, demonstrating the model's ability to model temporal alignment between audio and video.
	Audio Visual Interaction	Interactive communication with the model using audio-visual inputs, including task specification via audio.
	Audio Visual Dialogue	Conversational interaction with the model using audio-visual inputs, showcasing its capabilities in casual chat and assistant-like behavior.
Agent	Audio Function Call	Using audio input to perform function calls, enabling agent-like behaviors.
Downstream Task Fine-tuning	Omni Captioner	Introduction and capability demonstration of Qwen3-Omni-30B-A3B-Captioner, a downstream fine-tuned model based on Qwen3-Omni-30B-A3B-Instruct, illustrating the strong generalization ability of the Qwen3-Omni foundation model.

## QuickStart ### Model Description and Download Below is the description of all Qwen3-Omni models. Please select and download the model that fits your needs. | Model Name | Description | |------------------------------|-------------| | Qwen3-Omni-30B-A3B-Instruct | The Instruct model of Qwen3-Omni-30B-A3B, containing both thinker and talker, supporting audio, video, and text input, with audio and text output. For more information, please read the [Qwen3-Omni Technical Report](https://github.com/QwenLM/Qwen3-Omni/blob/main/assets/Qwen3_Omni.pdf). | | Qwen3-Omni-30B-A3B-Thinking | The Thinking model of Qwen3-Omni-30B-A3B, containing the thinker component, equipped with chain-of-thought reasoning, supporting audio, video, and text input, with text output. For more information, please read the [Qwen3-Omni Technical Report](https://github.com/QwenLM/Qwen3-Omni/blob/main/assets/Qwen3_Omni.pdf).| | Qwen3-Omni-30B-A3B-Captioner | A downstream audio fine-grained caption model fine-tuned from Qwen3-Omni-30B-A3B-Instruct, which produces detailed, low-hallucination captions for arbitrary audio inputs. It contains the thinker, supporting audio input and text output. For more information, you can refer to the model's [cookbook](https://github.com/QwenLM/Qwen3-Omni/blob/main/cookbooks/omni_captioner.ipynb). | During loading in Hugging Face Transformers or vLLM, model weights will be automatically downloaded based on the model name. However, if your runtime environment is not conducive to downloading weights during execution, you can refer to the following commands to manually download the model weights to a local directory: ```bash # Download through ModelScope (recommended for users in Mainland China) pip install -U modelscope modelscope download --model Qwen/Qwen3-Omni-30B-A3B-Instruct --local_dir ./Qwen3-Omni-30B-A3B-Instruct modelscope download --model Qwen/Qwen3-Omni-30B-A3B-Thinking --local_dir ./Qwen3-Omni-30B-A3B-Thinking modelscope download --model Qwen/Qwen3-Omni-30B-A3B-Captioner --local_dir ./Qwen3-Omni-30B-A3B-Captioner # Download through Hugging Face pip install -U "huggingface_hub[cli]" huggingface-cli download Qwen/Qwen3-Omni-30B-A3B-Instruct --local-dir ./Qwen3-Omni-30B-A3B-Instruct huggingface-cli download Qwen/Qwen3-Omni-30B-A3B-Thinking --local-dir ./Qwen3-Omni-30B-A3B-Thinking huggingface-cli download Qwen/Qwen3-Omni-30B-A3B-Captioner --local-dir ./Qwen3-Omni-30B-A3B-Captioner ``` ### Transformers Usage #### Installation The Hugging Face Transformers code for Qwen3-Omni has been successfully merged, but the PyPI package has not yet been released. Therefore, you need to install it from source using the following command. We strongly recommend that you **create a new Python environment** to avoid environment runtime issues. ```bash # If you already have transformers installed, please uninstall it first, or create a new Python environment # pip uninstall transformers pip install git+https://github.com/huggingface/transformers pip install accelerate ``` We offer a toolkit to help you handle various types of audio and visual input more conveniently, providing an API-like experience. This includes support for base64, URLs, and interleaved audio, images, and videos. You can install it using the following command and make sure your system has `ffmpeg` installed: ```bash pip install qwen-omni-utils -U ``` Additionally, we recommend using FlashAttention 2 when running with Hugging Face Transformers to reduce GPU memory usage. However, if you are primarily using [vLLM](#vllm-usage) for inference, this installation is not necessary, as vLLM includes FlashAttention 2 by default. ```bash pip install -U flash-attn --no-build-isolation ``` Also, you should have hardware that is compatible with FlashAttention 2. Read more about it in the official documentation of the [FlashAttention repository](https://github.com/Dao-AILab/flash-attention). FlashAttention 2 can only be used when a model is loaded in `torch.float16` or `torch.bfloat16`. #### Code Snippet Here is a code snippet to show you how to use Qwen3-Omni with `transformers` and `qwen_omni_utils`: ```python import soundfile as sf from transformers import Qwen3OmniMoeForConditionalGeneration, Qwen3OmniMoeProcessor from qwen_omni_utils import process_mm_info MODEL_PATH = "Qwen/Qwen3-Omni-30B-A3B-Instruct" # MODEL_PATH = "Qwen/Qwen3-Omni-30B-A3B-Thinking" model = Qwen3OmniMoeForConditionalGeneration.from_pretrained( MODEL_PATH, dtype="auto", device_map="auto", attn_implementation="flash_attention_2", ) processor = Qwen3OmniMoeProcessor.from_pretrained(MODEL_PATH) conversation = [ { "role": "user", "content": [ {"type": "image", "image": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-Omni/demo/cars.jpg"}, {"type": "audio", "audio": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-Omni/demo/cough.wav"}, {"type": "text", "text": "What can you see and hear? Answer in one short sentence."} ], }, ] # Set whether to use audio in video USE_AUDIO_IN_VIDEO = True # Preparation for inference text = processor.apply_chat_template(conversation, add_generation_prompt=True, tokenize=False) audios, images, videos = process_mm_info(conversation, use_audio_in_video=USE_AUDIO_IN_VIDEO) inputs = processor(text=text, audio=audios, images=images, videos=videos, return_tensors="pt", padding=True, use_audio_in_video=USE_AUDIO_IN_VIDEO) inputs = inputs.to(model.device).to(model.dtype) # Inference: Generation of the output text and audio text_ids, audio = model.generate(**inputs, speaker="Ethan", thinker_return_dict_in_generate=True, use_audio_in_video=USE_AUDIO_IN_VIDEO) text = processor.batch_decode(text_ids.sequences[:, inputs["input_ids"].shape[1] :], skip_special_tokens=True, clean_up_tokenization_spaces=False) print(text) if audio is not None: sf.write( "output.wav", audio.reshape(-1).detach().cpu().numpy(), samplerate=24000, ) ``` Here are some more advanced usage examples. You can expand the sections below to learn more.

Batch inference

The model can batch inputs composed of mixed samples of various types such as text, images, audio, and videos as input when `return_audio=False` is set. Here is an example. ```python from transformers import Qwen3OmniMoeForConditionalGeneration, Qwen3OmniMoeProcessor from qwen_omni_utils import process_mm_info MODEL_PATH = "Qwen/Qwen3-Omni-30B-A3B-Instruct" # MODEL_PATH = "Qwen/Qwen3-Omni-30B-A3B-Thinking" model = Qwen3OmniMoeForConditionalGeneration.from_pretrained( MODEL_PATH, dtype="auto", device_map="auto", attn_implementation="flash_attention_2", ) model.disable_talker() processor = Qwen3OmniMoeProcessor.from_pretrained(MODEL_PATH) # Conversation with image only conversation1 = [ { "role": "user", "content": [ {"type": "image", "image": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-Omni/demo/cars.jpg"}, {"type": "text", "text": "What can you see in this image? Answer in one sentence."}, ] } ] # Conversation with audio only conversation2 = [ { "role": "user", "content": [ {"type": "audio", "audio": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-Omni/demo/cough.wav"}, {"type": "text", "text": "What can you hear in this audio?"}, ] } ] # Conversation with pure text and system prompt conversation3 = [ { "role": "system", "content": [ {"type": "text", "text": "You are Qwen-Omni."} ], }, { "role": "user", "content": "Who are you?" } ] # Conversation with mixed media conversation4 = [ { "role": "user", "content": [ {"type": "image", "image": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-Omni/demo/cars.jpg"}, {"type": "audio", "audio": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-Omni/demo/cough.wav"}, {"type": "text", "text": "What can you see and hear? Answer in one sentence."} ], } ] # Combine messages for batch processing conversations = [conversation1, conversation2, conversation3, conversation4] # Set whether to use audio in video USE_AUDIO_IN_VIDEO = True # Preparation for batch inference text = processor.apply_chat_template(conversations, add_generation_prompt=True, tokenize=False) audios, images, videos = process_mm_info(conversations, use_audio_in_video=USE_AUDIO_IN_VIDEO) inputs = processor(text=text, audio=audios, images=images, videos=videos, return_tensors="pt", padding=True, use_audio_in_video=USE_AUDIO_IN_VIDEO) inputs = inputs.to(model.device).to(model.dtype) # Batch inference does not support returning audio text_ids, audio = model.generate(**inputs, return_audio=False, thinker_return_dict_in_generate=True, use_audio_in_video=USE_AUDIO_IN_VIDEO) text = processor.batch_decode(text_ids.sequences[:, inputs["input_ids"].shape[1] :], skip_special_tokens=True, clean_up_tokenization_spaces=False) print(text) ```

Use audio output or not

The model supports both text and audio outputs. If users do not need audio outputs, they can call `model.disable_talker()` after initializing the model. This option will save about `10GB` of GPU memory, but the `return_audio` option for the `generate` function will only allow `False`. ```python model = Qwen3OmniMoeForConditionalGeneration.from_pretrained( "Qwen/Qwen3-Omni-30B-A3B-Instruct", dtype="auto", device_map="auto", attn_implementation="flash_attention_2", ) model.disable_talker() ``` For a more flexible experience, we recommend that users decide whether to return audio when the `generate` function is called. If `return_audio` is set to `False`, the model will only return text outputs, resulting in faster text responses. ```python model = Qwen3OmniMoeForConditionalGeneration.from_pretrained( "Qwen/Qwen3-Omni-30B-A3B-Instruct", dtype="auto", device_map="auto", attn_implementation="flash_attention_2", ) ... text_ids, _ = model.generate(..., return_audio=False)```

Change voice type of output audio

Qwen3-Omni supports changing the voice of the output audio. The `"Qwen/Qwen3-Omni-30B-A3B-Instruct"` checkpoint supports three voice types as follows: | Voice Type | Gender | Description | |------------|--------|-------------| | Ethan | Male | A bright, upbeat voice with infectious energy and a warm, approachable vibe. | | Chelsie | Female | A honeyed, velvety voice that carries a gentle warmth and luminous clarity. | | Aiden | Male | A warm, laid-back American voice with a gentle, boyish charm. | Users can use the `speaker` parameter of the `generate` function to specify the voice type. By default, if `speaker` is not specified, the voice type is `Ethan`. ```python text_ids, audio = model.generate(..., speaker="Ethan") ``` ```python text_ids, audio = model.generate(..., speaker="Chelsie") ``` ```python text_ids, audio = model.generate(..., speaker="Aiden") ```

### vLLM Usage #### Installation We strongly recommend using vLLM for inference and deployment of the Qwen3-Omni series models. Since our code is currently in the pull request stage, and **audio output inference support for the Instruct model will be released in the near future**, you can follow the commands below to install vLLM from source. Please note that we recommend you **create a new Python environment** to avoid runtime environment conflicts and incompatibilities. For more details on compiling vLLM from source, please refer to the [vLLM official documentation](https://docs.vllm.ai/en/latest/getting_started/installation/gpu.html#set-up-using-python-only-build-without-compilation). ```bash git clone -b qwen3_omni https://github.com/wangxiongts/vllm.git cd vllm pip install -r requirements/build.txt pip install -r requirements/cuda.txt export VLLM_PRECOMPILED_WHEEL_LOCATION=https://wheels.vllm.ai/a5dd03c1ebc5e4f56f3c9d3dc0436e9c582c978f/vllm-0.9.2-cp38-abi3-manylinux1_x86_64.whl VLLM_USE_PRECOMPILED=1 pip install -e . -v --no-build-isolation # If you meet an "Undefined symbol" error while using VLLM_USE_PRECOMPILED=1, please use "pip install -e . -v" to build from source. # Install the Transformers pip install git+https://github.com/huggingface/transformers pip install accelerate pip install qwen-omni-utils -U pip install -U flash-attn --no-build-isolation ``` #### Inference You can use the following code for vLLM inference. The `limit_mm_per_prompt` parameter specifies the maximum number of each modality's data allowed per message. Since vLLM needs to pre-allocate GPU memory, larger values will require more GPU memory; if OOM issues occur, try reducing this value. Setting `tensor_parallel_size` greater than one enables multi-GPU parallel inference, improving concurrency and throughput. In addition, `max_num_seqs` indicates the number of sequences that vLLM processes in parallel during each inference step. A larger value requires more GPU memory but enables higher batch inference speed. For more details, please refer to the [vLLM official documentation](https://docs.vllm.ai/en/latest/api/vllm/index.html#vllm.LLM). Below is a simple example of how to run Qwen3-Omni with vLLM: ```python import os import torch from vllm import LLM, SamplingParams from transformers import Qwen3OmniMoeProcessor from qwen_omni_utils import process_mm_info if __name__ == '__main__': # vLLM engine v1 not supported yet os.environ['VLLM_USE_V1'] = '0' MODEL_PATH = "Qwen/Qwen3-Omni-30B-A3B-Instruct" # MODEL_PATH = "Qwen/Qwen3-Omni-30B-A3B-Thinking" llm = LLM( model=MODEL_PATH, trust_remote_code=True, gpu_memory_utilization=0.95, tensor_parallel_size=torch.cuda.device_count(), limit_mm_per_prompt={'image': 3, 'video': 3, 'audio': 3}, max_num_seqs=8, max_model_len=32768, seed=1234, ) sampling_params = SamplingParams( temperature=0.6, top_p=0.95, top_k=20, max_tokens=16384, ) processor = Qwen3OmniMoeProcessor.from_pretrained(MODEL_PATH) messages = [ { "role": "user", "content": [ {"type": "video", "video": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-Omni/demo/draw.mp4"} ], } ] text = processor.apply_chat_template( messages, tokenize=False, add_generation_prompt=True, ) audios, images, videos = process_mm_info(messages, use_audio_in_video=True) inputs = { 'prompt': text, 'multi_modal_data': {}, "mm_processor_kwargs": { "use_audio_in_video": True, }, } if images is not None: inputs['multi_modal_data']['image'] = images if videos is not None: inputs['multi_modal_data']['video'] = videos if audios is not None: inputs['multi_modal_data']['audio'] = audios outputs = llm.generate([inputs], sampling_params=sampling_params) print(outputs[0].outputs[0].text) ``` Here are some more advanced usage examples. You can expand the sections below to learn more.

Batch inference

Using vLLM enables fast batch inference, which can help you efficiently process large volumes of data or conduct benchmarking. Refer to the following code example: ```python import os import torch from vllm import LLM, SamplingParams from transformers import Qwen3OmniMoeProcessor from qwen_omni_utils import process_mm_info def build_input(processor, messages, use_audio_in_video): text = processor.apply_chat_template( messages, tokenize=False, add_generation_prompt=True, ) audios, images, videos = process_mm_info(messages, use_audio_in_video=use_audio_in_video) inputs = { 'prompt': text, 'multi_modal_data': {}, "mm_processor_kwargs": { "use_audio_in_video": use_audio_in_video, }, } if images is not None: inputs['multi_modal_data']['image'] = images if videos is not None: inputs['multi_modal_data']['video'] = videos if audios is not None: inputs['multi_modal_data']['audio'] = audios return inputs if __name__ == '__main__': # vLLM engine v1 not supported yet os.environ['VLLM_USE_V1'] = '0' MODEL_PATH = "Qwen/Qwen3-Omni-30B-A3B-Instruct" # MODEL_PATH = "Qwen/Qwen3-Omni-30B-A3B-Thinking" llm = LLM( model=MODEL_PATH, trust_remote_code=True, gpu_memory_utilization=0.95, tensor_parallel_size=torch.cuda.device_count(), limit_mm_per_prompt={'image': 3, 'video': 3, 'audio': 3}, max_num_seqs=8, max_model_len=32768, seed=1234, ) sampling_params = SamplingParams( temperature=0.6, top_p=0.95, top_k=20, max_tokens=16384, ) processor = Qwen3OmniMoeProcessor.from_pretrained(MODEL_PATH) # Conversation with image only conversation1 = [ { "role": "user", "content": [ {"type": "image", "image": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-Omni/demo/cars.jpg"}, {"type": "text", "text": "What can you see in this image? Answer in one sentence."}, ] } ] # Conversation with audio only conversation2 = [ { "role": "user", "content": [ {"type": "audio", "audio": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-Omni/demo/cough.wav"}, {"type": "text", "text": "What can you hear in this audio?"}, ] } ] # Conversation with pure text and system prompt conversation3 = [ { "role": "system", "content": [ {"type": "text", "text": "You are Qwen-Omni."} ], }, { "role": "user", "content": "Who are you? Answer in one sentence." } ] # Conversation with mixed media conversation4 = [ { "role": "user", "content": [ {"type": "image", "image": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-Omni/demo/cars.jpg"}, {"type": "audio", "audio": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-Omni/cookbook/asr_fr.wav"}, {"type": "text", "text": "What can you see and hear? Answer in one sentence."} ], } ] USE_AUDIO_IN_VIDEO = True # Combine messages for batch processing conversations = [conversation1, conversation2, conversation3, conversation4] inputs = [build_input(processor, messages, USE_AUDIO_IN_VIDEO) for messages in conversations] outputs = llm.generate(inputs, sampling_params=sampling_params) result = [outputs[i].outputs[0].text for i in range(len(outputs))] print(result) ```

vLLM Serve Usage

vLLM serve for Qwen3-Omni currently only supports the thinker model. The `use_audio_in_video` parameter is not available in vLLM serve; you can handle this by separately passing video and audio inputs for processing. You can start vLLM serve through the following command: ```bash # Qwen3-Omni-30B-A3B-Instruct for single GPU vllm serve Qwen/Qwen3-Omni-30B-A3B-Instruct --port 8901 --host 127.0.0.1 --dtype bfloat16 --max-model-len 32768 --allowed-local-media-path / -tp 1 # Qwen3-Omni-30B-A3B-Instruct for multi-GPU (example on 4 GPUs) vllm serve Qwen/Qwen3-Omni-30B-A3B-Instruct --port 8901 --host 127.0.0.1 --dtype bfloat16 --max-model-len 65536 --allowed-local-media-path / -tp 4 # Qwen/Qwen3-Omni-30B-A3B-Thinking for single GPU vllm serve Qwen/Qwen3-Omni-30B-A3B-Thinking --port 8901 --host 127.0.0.1 --dtype bfloat16 --max-model-len 32768 --allowed-local-media-path / -tp 1 # Qwen/Qwen3-Omni-30B-A3B-Thinking for multi-GPU (example on 4 GPUs) vllm serve Qwen/Qwen3-Omni-30B-A3B-Thinking --port 8901 --host 127.0.0.1 --dtype bfloat16 --max-model-len 65536 --allowed-local-media-path / -tp 4 ``` Then you can use the chat API as below (via curl, for example): ```bash curl http://localhost:8901/v1/chat/completions \ -H "Content-Type: application/json" \ -d '{ "messages": [ {"role": "system", "content": "You are a helpful assistant."}, {"role": "user", "content": [ {"type": "image_url", "image_url": {"url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-Omni/demo/cars.jpg"}}, {"type": "audio_url", "audio_url": {"url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-Omni/demo/cough.wav"}}, {"type": "text", "text": "What can you see and hear? Answer in one sentence."} ]} ] }' ```

### Usage Tips (Recommended Reading) #### Minimum GPU memory requirements | Model | Precision | 15s Video | 30s Video | 60s Video | 120s Video | |------------------------------|-----------| --------- | --------- | --------- | --------- | | Qwen3-Omni-30B-A3B-Instruct | BF16 | 78.85 GB | 88.52 GB | 107.74 GB | 144.81 GB | | Qwen3-Omni-30B-A3B-Thinking | BF16 | 68.74 GB | 77.79 GB | 95.76 GB | 131.65 GB | **Note**: The table above presents the theoretical minimum memory requirements for inference with `transformers` and `BF16` precision, tested with `attn_implementation="flash_attention_2"`. The Instruct model includes both the **thinker** and **talker** components, whereas the Thinking model includes only the **thinker** part. #### Prompt for Audio-Visual Interaction When using Qwen3-Omni for audio-visual multimodal interaction, where the input consists of a video and its corresponding audio (with the audio serving as a query), we recommend using the **following system prompt**. This setup helps the model maintain high reasoning capability while better assuming interactive roles such as a smart assistant. Additionally, the text generated by the thinker will be more readable, with a natural, conversational tone and without complex formatting that is difficult to vocalize, leading to more stable and fluent audio output from the talker. You can customize the `user_system_prompt` field in the system prompt to include character settings or other role-specific descriptions as needed. ``` user_system_prompt = "You are Qwen-Omni, a smart voice assistant created by Alibaba Qwen." message = { "role": "system", "content": [ {"type": "text", "text": f"{user_system_prompt} You are a virtual voice assistant with no gender or age.\nYou are communicating with the user.\nIn user messages, “I/me/my/we/our” refer to the user and “you/your” refer to the assistant. In your replies, address the user as “you/your” and yourself as “I/me/my”; never mirror the user’s pronouns—always shift perspective. Keep original pronouns only in direct quotes; if a reference is unclear, ask a brief clarifying question.\nInteract with users using short(no more than 50 words), brief, straightforward language, maintaining a natural tone.\nNever use formal phrasing, mechanical expressions, bullet points, overly structured language. \nYour output must consist only of the spoken content you want the user to hear. \nDo not include any descriptions of actions, emotions, sounds, or voice changes. \nDo not use asterisks, brackets, parentheses, or any other symbols to indicate tone or actions. \nYou must answer users' audio or text questions, do not directly describe the video content. \nYou should communicate in the same language strictly as the user unless they request otherwise.\nWhen you are uncertain (e.g., you can't see/hear clearly, don't understand, or the user makes a comment rather than asking a question), use appropriate questions to guide the user to continue the conversation.\nKeep replies concise and conversational, as if talking face-to-face."} ] } ``` #### Best Practices for the Thinking Model The `Qwen3-Omni-30B-A3B-Thinking` model is primarily designed for understanding and interacting with multimodal inputs, including text, audio, image, and video. To achieve optimal performance, we recommend that users include an explicit textual instruction or task description in each round of dialogue alongside the multimodal input. This helps clarify the intent and significantly enhances the model's ability to leverage its reasoning capabilities. For example: ```python messages = [ { "role": "user", "content": [ {"type": "audio", "audio": "/path/to/audio.wav"}, {"type": "image", "image": "/path/to/image.png"}, {"type": "video", "video": "/path/to/video.mp4"}, {"type": "text", "text": "Analyze this audio, image, and video together."}, ], } ] ``` #### Use audio in video In multimodal interaction, user-provided videos are often accompanied by audio (such as spoken questions or sounds from events in the video). This information helps the model provide a better interactive experience. We provide the following options for users to decide whether to use the audio from a video. ```python # In data preprocessing audios, images, videos = process_mm_info(messages, use_audio_in_video=True) ``` ```python # For Transformers text = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=False) inputs = processor(text=text, audio=audios, images=images, videos=videos, return_tensors="pt", padding=True, use_audio_in_video=True) text_ids, audio = model.generate(..., use_audio_in_video=True) # For vLLM text = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=False) inputs = { 'prompt': text, 'multi_modal_data': {}, "mm_processor_kwargs": { "use_audio_in_video": True, }, } ``` It is worth noting that during a multi-round conversation, the `use_audio_in_video` parameter must be set consistently across these steps; otherwise, unexpected results may occur. ## Evaluation ### Performance of Qwen3-Omni Qwen3-Omni maintains state-of-the-art performance on text and visual modalities without degradation relative to same-size single-model Qwen counterparts. Across 36 audio and audio-visual benchmarks, it achieves open-source SOTA on 32 and sets the SOTA on 22, outperforming strong closed-source systems such as Gemini 2.5 Pro and GPT-4o.

Text -> Text

		GPT-4o-0327	Qwen3-235B-A22B Non Thinking	Qwen3-30B-A3B-Instruct-2507	Qwen3-Omni-30B-A3B-Instruct	Qwen3-Omni-Flash-Instruct
General Tasks	MMLU-Redux	91.3	89.2	89.3	86.6	86.8
General Tasks	GPQA	66.9	62.9	70.4	69.6	69.7
Reasoning	AIME25	26.7	24.7	61.3	65.0	65.9
Reasoning	ZebraLogic	52.6	37.7	90.0	76.0	76.1
Code	MultiPL-E	82.7	79.3	83.8	81.4	81.5
Alignment Tasks	IFEval	83.9	83.2	84.7	81.0	81.7
	Creative Writing v3	84.9	80.4	86.0	80.6	81.8
	WritingBench	75.5	77.0	85.5	82.6	83.0
Agent	BFCL-v3	66.5	68.0	65.1	64.4	65.0
Multilingual Tasks	MultiIF	70.4	70.2	67.9	64.0	64.7
Multilingual Tasks	PolyMATH	25.5	27.0	43.1	37.9	39.3

		Gemini-2.5-Flash Thinking	Qwen3-235B-A22B Thinking	Qwen3-30B-A3B-Thinking-2507	Qwen3-Omni-30B-A3B-Thinking	Qwen3-Omni-Flash-Thinking
General Tasks	MMLU-Redux	92.1	92.7	91.4	88.8	89.7
General Tasks	GPQA	82.8	71.1	73.4	73.1	73.1
Reasoning	AIME25	72.0	81.5	85.0	73.7	74.0
Reasoning	LiveBench 20241125	74.3	77.1	76.8	71.8	70.3
Code	MultiPL-E	84.5	79.9	81.3	80.6	81.0
Alignment Tasks	IFEval	89.8	83.4	88.9	85.1	85.2
	Arena-Hard v2	56.7	61.5	56.0	55.1	57.8
	Creative Writing v3	85.0	84.6	84.4	82.5	83.6
	WritingBench	83.9	80.3	85.0	85.5	85.9
Agent	BFCL-v3	68.6	70.8	72.4	63.2	64.5
Multilingual Tasks	MultiIF	74.4	71.9	76.4	72.9	73.2
Multilingual Tasks	PolyMATH	49.8	54.7	52.6	47.1	48.7

Audio -> Text

	Seed-ASR	Voxtral-Mini	Voxtral-Small	GPT-4o-Transcribe	Gemini-2.5-Pro	Qwen2.5-Omni	Qwen3-Omni-30B-A3B-Instruct	Qwen3-Omni-Flash-Instruct
EN & ZH ASR (wer)
Wenetspeech net \| meeting	4.66 \| 5.69	24.30 \| 31.53	20.33 \| 26.08	15.30 \| 32.27	14.43 \| 13.47	5.91 \| 7.65	4.69 \| 5.89	4.62 \| 5.75
Librispeech clean \| other	1.58 \| 2.84	1.88 \| 4.12	1.56 \| 3.30	1.39 \| 3.75	2.89 \| 3.56	1.74 \| 3.45	1.22 \| 2.48	1.27 \| 2.44
CV15-en	-	9.47	7.79	10.01	9.89	7.61	6.05	5.94
CV15-zh	-	24.67	19.30	9.84	8.00	5.13	4.31	4.28
Fleurs-en	3.40	3.96	3.77	3.32	2.94	3.77	2.72	2.74
Fleurs-zh	2.69	12.22	7.98	2.44	2.71	2.54	2.20	2.19
Multilingual ASR (wer)
Fleurs-avg (19 lang)	-	15.67	8.09	4.48	5.55	14.04	5.33	5.31
Lyric ASR (wer)
MIR-1K (vocal-only)	6.45	23.33	18.73	11.87	9.85	8.15	5.90	5.85
Opencpop-test	2.98	31.01	16.06	7.93	6.49	2.84	1.54	2.02
S2TT (BLEU)
Fleurs-en2xx	-	30.35	37.85	-	39.25	29.22	37.50	36.22
Fleurs-xx2en	-	27.54	32.81	-	35.41	28.61	31.08	30.71
Fleurs-zh2xx	-	17.03	22.05	-	26.63	17.97	25.17	25.10
Fleurs-xx2zh	-	28.75	34.82	-	37.50	27.68	33.13	31.19

	GPT-4o-Audio	Gemini-2.5-Flash	Gemini-2.5-Pro	Qwen2.5-Omni	Qwen3-Omni-30B-A3B-Instruct	Qwen3-Omni-30B-A3B-Thinking	Qwen3-Omni-Flash-Instruct	Qwen3-Omni-Flash-Thinking
VoiceBench
AlpacaEval	95.6	96.1	94.3	89.9	94.8	96.4	95.4	96.8
CommonEval	89.8	88.3	88.4	76.7	90.8	90.5	91.0	90.9
WildVoice	91.6	92.1	93.4	77.7	91.6	90.5	92.3	90.9
SD-QA	75.5	84.5	90.1	56.4	76.9	78.1	76.8	78.5
MMSU	80.3	66.1	71.1	61.7	68.1	83.0	68.4	84.3
OpenBookQA	89.2	56.9	92.3	80.9	89.7	94.3	91.4	95.0
BBH	84.1	83.9	92.6	66.7	80.4	88.9	80.6	89.6
IFEval	76.0	83.8	85.7	53.5	77.8	80.6	75.2	80.8
AdvBench	98.7	98.9	98.1	99.2	99.3	97.2	99.4	98.9
Overall	86.8	83.4	89.6	73.6	85.5	88.8	85.6	89.5
Audio Reasoning
MMAU-v05.15.25	62.5	71.8	77.4	65.5	77.5	75.4	77.6	76.5
MMSU	56.4	70.2	77.7	62.6	69.0	70.2	69.1	71.3

	Best Specialist Models	GPT-4o-Audio	Gemini-2.5-Pro	Qwen2.5-Omni	Qwen3-Omni-30B-A3B-Instruct	Qwen3-Omni-Flash-Instruct
RUL-MuchoMusic	47.6 (Audio Flamingo 3)	36.1	49.4	47.3	52.0	52.1
GTZAN Acc.	87.9 (CLaMP 3)	76.5	81.0	81.7	93.0	93.1
MTG Genre Micro F1	35.8 (MuQ-MuLan)	25.3	32.6	32.5	39.0	39.5
MTG Mood/Theme Micro F1	10.9 (MuQ-MuLan)	11.3	14.1	8.9	21.0	21.7
MTG Instrument Micro F1	39.8 (MuQ-MuLan)	34.2	33.0	22.6	40.5	40.7
MTG Top50 Micro F1	33.2 (MuQ-MuLan)	25.0	26.1	21.6	36.7	36.9
MagnaTagATune Micro F1	41.6 (MuQ)	29.2	28.1	30.1	44.3	46.8

Vision -> Text

Datasets	GPT4-o	Gemini-2.0-Flash	Qwen2.5-VL 72B	Qwen3-Omni-30B-A3B -Instruct	Qwen3-Omni-Flash -Instruct
General Visual Question Answering
MMStar	64.7	71.4	70.8	68.5	69.3
HallusionBench	55.0	56.3	55.2	59.7	58.5
MM-MT-Bench	7.7	6.7	7.6	7.4	7.6
Math & STEM
MMMU_val	69.1	71.3	70.2	69.1	69.8
MMMU_pro	51.9	56.1	51.1	57.0	57.6
MathVista_mini	63.8	71.4	74.8	75.9	77.4
MathVision_full	30.4	48.6	38.1	56.3	58.3
Documentation Understanding
AI2D	84.6	86.7	88.7	85.2	86.4
ChartQA_test	86.7	64.6	89.5	86.8	87.1
Counting
CountBench	87.9	91.2	93.6	90.0	90.0
Video Understanding
Video-MME	71.9	72.4	73.3	70.5	71.4
LVBench	30.8	57.9	47.3	50.2	51.1
MLVU	64.6	71.0	74.6	75.2	75.7

Datasets	Gemini-2.5-flash-thinking	InternVL-3.5-241B-A28B	Qwen3-Omni-30B-A3B-Thinking	Qwen3-Omni-Flash-Thinking
General Visual Question Answering
MMStar	75.5	77.9	74.9	75.5
HallusionBench	61.1	57.3	62.8	63.4
MM-MT-Bench	7.8	–	8.0	8.0
Math & STEM
MMMU_val	76.9	77.7	75.6	75.0
MMMU_pro	65.8	–	60.5	60.8
MathVista_mini	77.6	82.7	80.0	81.2
MathVision_full	62.3	63.9	62.9	63.8
Documentation Understanding
AI2D_test	88.6	87.3	86.1	86.8
ChartQA_test	–	88.0	89.5	89.3
Counting
CountBench	88.6	–	88.6	92.5
Video Understanding
Video-MME	79.6	72.9	69.7	69.8
LVBench	64.5	–	49.0	49.5
MLVU	82.1	78.2	72.9	73.9

AudioVisual -> Text

Datasets	Previous Open-source SoTA	Gemini-2.5-Flash	Qwen2.5-Omni	Qwen3-Omni-30B-A3B-Instruct	Qwen3-Omni-Flash-Instruct
WorldSense	47.1	50.9	45.4	54.0	54.1

Datasets	Previous Open-source SoTA	Gemini-2.5-Flash-Thinking	Qwen3-Omni-30B-A3B-Thinking	Qwen3-Omni-Flash-Thinking
DailyOmni	69.8	72.7	75.8	76.2
VideoHolmes	55.6	49.5	57.3	57.3

Zero-shot Speech Generation

Datasets	Model	Performance
	Content Consistency
SEED test-zh \| test-en	Seed-TTS_ICL	1.11 \| 2.24
	Seed-TTS_RL	1.00 \| 1.94
	MaskGCT	2.27 \| 2.62
	E2 TTS	1.97 \| 2.19
	F5-TTS	1.56 \| 1.83
	Spark TTS	1.20 \| 1.98
	CosyVoice 2	1.45 \| 2.57
	CosyVoice 3	0.71 \| 1.45
	Qwen2.5-Omni-7B	1.42 \| 2.33
	Qwen3-Omni-30B-A3B	1.07 \| 1.39

Multilingual Speech Generation

Language	Content Consistency			Speaker Similarity
Language	Qwen3-Omni-30B-A3B	MiniMax	ElevenLabs	Qwen3-Omni-30B-A3B	MiniMax	ElevenLabs
Chinese	0.716	2.252	16.026	0.772	0.780	0.677
English	1.069	2.164	2.339	0.773	0.756	0.613
German	0.777	1.906	0.572	0.738	0.733	0.614
Italian	1.067	1.543	1.743	0.742	0.699	0.579
Portuguese	1.872	1.877	1.331	0.770	0.805	0.711
Spanish	1.765	1.029	1.084	0.744	0.762	0.615
Japanese	3.631	3.519	10.646	0.763	0.776	0.738
Korean	1.670	1.747	1.865	0.778	0.776	0.700
French	2.505	4.099	5.216	0.689	0.628	0.535
Russian	3.986	4.281	3.878	0.759	0.761	0.676

Cross-Lingual Speech Generation

Language	Qwen3-Omni-30B-A3B	CosyVoice3	CosyVoice2
en-to-zh	5.37	5.09	13.5
ja-to-zh	3.32	3.05	48.1
ko-to-zh	0.99	1.06	7.70
zh-to-en	2.76	2.98	6.47
ja-to-en	3.31	4.20	17.1
ko-to-en	3.34	4.19	11.2
zh-to-ja	8.29	7.08	13.1
en-to-ja	7.53	6.80	14.9
ko-to-ja	4.24	3.93	5.86
zh-to-ko	5.13	14.4	24.8
en-to-ko	4.96	5.87	21.9
ja-to-ko	6.23	7.92	21.5

### Setting for Evaluation * **Decoding Strategy**: For the Qwen3-Omni series across all evaluation benchmarks, `Instruct` models use greedy decoding during generation without sampling. For `Thinking` models, the decoding parameters should be taken from the `generation_config.json` file in the checkpoint. * **Benchmark-Specific Formatting**: For the majority of evaluation benchmarks, they come with their own ChatML formatting to embed the question or prompt. It should be noted that all video data are set to `fps=2` during evaluation. * **Default Prompts**: For tasks in certain benchmarks that do not include a prompt, we use the following prompt settings: | Task Type | Prompt | | :--- | :--- | | Auto Speech Recognition (ASR) for Chinese | 请将这段中文语音转换为纯文本。 | | Auto Speech Recognition (ASR) for Other languages | Transcribe the audio into text. | | Speech-to-Text Translation (S2TT) | Listen to the provided speech and produce a translation in text. | | Song Lyrics Recognition | Transcribe the song lyrics into text without any punctuation, separate lines with line breaks, and output only the lyrics without additional explanations. | * **System Prompt**: No `system prompt` should be set for any evaluation benchmark. * **Input Sequence**: The question or prompt should be input as user text. Unless otherwise specified by the benchmark, the text should come **after** multimodal data in the sequence. For example: ```python messages = [ { "role": "user", "content": [ {"type": "audio", "audio": "/path/to/audio.wav"}, {"type": "image", "image": "/path/to/image.png"}, {"type": "video", "video": "/path/to/video.mp4"}, {"type": "text", "text": "Describe the audio, image and video."}, ], }, ] ```