I have found that the standard IMatrix does not perform very well at low bit quantiztion and for MOE models. So I am using llama.cpp --tensor-type to bump up selected layers. See Layer bumping with llama.cpp

This does create larger model files but increases precision for a given model size.

Please provide feedback on how you find this method performs

Choosing the Right Model Format

Selecting the correct model format depends on your hardware capabilities and memory constraints.

BF16 (Brain Float 16) – Use if BF16 acceleration is available

A 16-bit floating-point format designed for faster computation while retaining good precision.
Provides similar dynamic range as FP32 but with lower memory usage.
Recommended if your hardware supports BF16 acceleration (check your device's specs).
Ideal for high-performance inference with reduced memory footprint compared to FP32.

📌 Use BF16 if:
✔ Your hardware has native BF16 support (e.g., newer GPUs, TPUs).
✔ You want higher precision while saving memory.
✔ You plan to requantize the model into another format.

📌 Avoid BF16 if:
❌ Your hardware does not support BF16 (it may fall back to FP32 and run slower).
❌ You need compatibility with older devices that lack BF16 optimization.

F16 (Float 16) – More widely supported than BF16

A 16-bit floating-point high precision but with less of range of values than BF16.
Works on most devices with FP16 acceleration support (including many GPUs and some CPUs).
Slightly lower numerical precision than BF16 but generally sufficient for inference.

📌 Use F16 if:
✔ Your hardware supports FP16 but not BF16.
✔ You need a balance between speed, memory usage, and accuracy.
✔ You are running on a GPU or another device optimized for FP16 computations.

📌 Avoid F16 if:
❌ Your device lacks native FP16 support (it may run slower than expected).
❌ You have memory limitations.

Hybrid Precision Models (e.g., `bf16_q8_0`, `f16_q4_K`) – Best of Both Worlds

These formats selectively quantize non-essential layers while keeping key layers in full precision (e.g., attention and output layers).

Named like bf16_q8_0 (meaning full-precision BF16 core layers + quantized Q8_0 other layers).
Strike a balance between memory efficiency and accuracy, improving over fully quantized models without requiring the full memory of BF16/F16.

📌 Use Hybrid Models if:
✔ You need better accuracy than quant-only models but can’t afford full BF16/F16 everywhere.
✔ Your device supports mixed-precision inference.
✔ You want to optimize trade-offs for production-grade models on constrained hardware.

📌 Avoid Hybrid Models if:
❌ Your target device doesn’t support mixed or full-precision acceleration.
❌ You are operating under ultra-strict memory limits (in which case use fully quantized formats).

Quantized Models (Q4_K, Q6_K, Q8, etc.) – For CPU & Low-VRAM Inference

Quantization reduces model size and memory usage while maintaining as much accuracy as possible.

Lower-bit models (Q4_K) → Best for minimal memory usage, may have lower precision.
Higher-bit models (Q6_K, Q8_0) → Better accuracy, requires more memory.

📌 Use Quantized Models if:
✔ You are running inference on a CPU and need an optimized model.
✔ Your device has low VRAM and cannot load full-precision models.
✔ You want to reduce memory footprint while keeping reasonable accuracy.

📌 Avoid Quantized Models if:
❌ You need maximum accuracy (full-precision models are better for this).
❌ Your hardware has enough VRAM for higher-precision formats (BF16/F16).

Very Low-Bit Quantization (IQ3_XS, IQ3_S, IQ3_M, Q4_K, Q4_0)

These models are optimized for very high memory efficiency, making them ideal for low-power devices or large-scale deployments where memory is a critical constraint.

IQ3_XS: Ultra-low-bit quantization (3-bit) with very high memory efficiency.
- Use case: Best for ultra-low-memory devices where even Q4_K is too large.
- Trade-off: Lower accuracy compared to higher-bit quantizations.
IQ3_S: Small block size for maximum memory efficiency.
- Use case: Best for low-memory devices where IQ3_XS is too aggressive.
IQ3_M: Medium block size for better accuracy than IQ3_S.
- Use case: Suitable for low-memory devices where IQ3_S is too limiting.
Q4_K: 4-bit quantization with block-wise optimization for better accuracy.
- Use case: Best for low-memory devices where Q6_K is too large.
Q4_0: Pure 4-bit quantization, optimized for ARM devices.
- Use case: Best for ARM-based devices or low-memory environments.

Ultra Low-Bit Quantization (IQ1_S IQ1_M IQ2_S IQ2_M IQ2_XS IQ2_XSS)

*Ultra-low-bit quantization (1 2-bit) with extreme memory efficiency.
- Use case: Best for cases were you have to fit the model into very constrained memory
- Trade-off: Very Low Accuracy. May not function as expected. Please test fully before using.

Summary Table: Model Format Selection

Model Format	Precision	Memory Usage	Device Requirements	Best Use Case
BF16	Very High	High	BF16-supported GPU/CPU	High-speed inference with reduced memory
F16	High	High	FP16-supported GPU/CPU	Inference when BF16 isn’t available
Q4_K	Medium-Low	Low	CPU or Low-VRAM devices	Memory-constrained inference
Q6_K	Medium	Moderate	CPU with more memory	Better accuracy with quantization
Q8_0	High	Moderate	GPU/CPU with moderate VRAM	Highest accuracy among quantized models
IQ3_XS	Low	Very Low	Ultra-low-memory devices	Max memory efficiency, low accuracy
IQ3_S	Low	Very Low	Low-memory devices	Slightly more usable than IQ3_XS
IQ3_M	Low-Medium	Low	Low-memory devices	Better accuracy than IQ3_S
Q4_0	Low	Low	ARM-based/embedded devices	Llama.cpp automatically optimizes for ARM inference
*Ultra Low-Bit (IQ1/2_)**	Very Low	Extremely Low	Tiny edge/embedded devices	Fit models in extremely tight memory; low accuracy
Hybrid (e.g., `bf16_q8_0`)	Medium–High	Medium	Mixed-precision capable hardware	Balanced performance and memory, near-FP accuracy in critical layers

Qwen2.5-Omni

Overview

Introduction

Qwen2.5-Omni is an end-to-end multimodal model designed to perceive diverse modalities, including text, images, audio, and video, while simultaneously generating text and natural speech responses in a streaming manner.

Key Features

Omni and Novel Architecture: We propose Thinker-Talker architecture, an end-to-end multimodal model designed to perceive diverse modalities, including text, images, audio, and video, while simultaneously generating text and natural speech responses in a streaming manner. We propose a novel position embedding, named TMRoPE (Time-aligned Multimodal RoPE), to synchronize the timestamps of video inputs with audio.
Real-Time Voice and Video Chat: Architecture designed for fully real-time interactions, supporting chunked input and immediate output.
Natural and Robust Speech Generation: Surpassing many existing streaming and non-streaming alternatives, demonstrating superior robustness and naturalness in speech generation.
Strong Performance Across Modalities: Exhibiting exceptional performance across all modalities when benchmarked against similarly sized single-modality models. Qwen2.5-Omni outperforms the similarly sized Qwen2-Audio in audio capabilities and achieves comparable performance to Qwen2.5-VL-7B.
Excellent End-to-End Speech Instruction Following: Qwen2.5-Omni shows performance in end-to-end speech instruction following that rivals its effectiveness with text inputs, evidenced by benchmarks such as MMLU and GSM8K.

Model Architecture

Performance

We conducted a comprehensive evaluation of Qwen2.5-Omni, which demonstrates strong performance across all modalities when compared to similarly sized single-modality models and closed-source models like Qwen2.5-VL-7B, Qwen2-Audio, and Gemini-1.5-pro. In tasks requiring the integration of multiple modalities, such as OmniBench, Qwen2.5-Omni achieves state-of-the-art performance. Furthermore, in single-modality tasks, it excels in areas including speech recognition (Common Voice), translation (CoVoST2), audio understanding (MMAU), image reasoning (MMMU, MMStar), video understanding (MVBench), and speech generation (Seed-tts-eval and subjective naturalness).

Multimodality -> Text

Datasets	Model	Performance
OmniBench Speech \| Sound Event \| Music \| Avg	Gemini-1.5-Pro	42.67%\|42.26%\|46.23%\|42.91%
MIO-Instruct	36.96%\|33.58%\|11.32%\|33.80%
AnyGPT (7B)	17.77%\|20.75%\|13.21%\|18.04%
video-SALMONN	34.11%\|31.70%\|56.60%\|35.64%
UnifiedIO2-xlarge	39.56%\|36.98%\|29.25%\|38.00%
UnifiedIO2-xxlarge	34.24%\|36.98%\|24.53%\|33.98%
MiniCPM-o	-\|-\|-\|40.50%
Baichuan-Omni-1.5	-\|-\|-\|42.90%
Qwen2.5-Omni-3B	52.14%\|52.08%\|52.83%\|52.19%
Qwen2.5-Omni-7B	55.25%\|60.00%\|52.83%\|56.13%

Audio -> Text

Datasets	Model	Performance
ASR
Librispeech dev-clean \| dev other \| test-clean \| test-other	SALMONN	-\|-\|2.1\|4.9
SpeechVerse	-\|-\|2.1\|4.4
Whisper-large-v3	-\|-\|1.8\|3.6
Llama-3-8B	-\|-\|-\|3.4
Llama-3-70B	-\|-\|-\|3.1
Seed-ASR-Multilingual	-\|-\|1.6\|2.8
MiniCPM-o	-\|-\|1.7\|-
MinMo	-\|-\|1.7\|3.9
Qwen-Audio	1.8\|4.0\|2.0\|4.2
Qwen2-Audio	1.3\|3.4\|1.6\|3.6
Qwen2.5-Omni-3B	2.0\|4.1\|2.2\|4.5
Qwen2.5-Omni-7B	1.6\|3.5\|1.8\|3.4
Common Voice 15 en \| zh \| yue \| fr	Whisper-large-v3	9.3\|12.8\|10.9\|10.8
MinMo	7.9\|6.3\|6.4\|8.5
Qwen2-Audio	8.6\|6.9\|5.9\|9.6
Qwen2.5-Omni-3B	9.1\|6.0\|11.6\|9.6
Qwen2.5-Omni-7B	7.6\|5.2\|7.3\|7.5
Fleurs zh \| en	Whisper-large-v3	7.7\|4.1
Seed-ASR-Multilingual	-\|3.4
Megrez-3B-Omni	10.8\|-
MiniCPM-o	4.4\|-
MinMo	3.0\|3.8
Qwen2-Audio	7.5\|-
Qwen2.5-Omni-3B	3.2\|5.4
Qwen2.5-Omni-7B	3.0\|4.1
Wenetspeech test-net \| test-meeting	Seed-ASR-Chinese	4.7\|5.7
Megrez-3B-Omni	-\|16.4
MiniCPM-o	6.9\|-
MinMo	6.8\|7.4
Qwen2.5-Omni-3B	6.3\|8.1
Qwen2.5-Omni-7B	5.9\|7.7
Voxpopuli-V1.0-en	Llama-3-8B	6.2
Llama-3-70B	5.7
Qwen2.5-Omni-3B	6.6
Qwen2.5-Omni-7B	5.8
S2TT
CoVoST2 en-de \| de-en \| en-zh \| zh-en	SALMONN	18.6\|-\|33.1\|-
SpeechLLaMA	-\|27.1\|-\|12.3
BLSP	14.1\|-\|-\|-
MiniCPM-o	-\|-\|48.2\|27.2
MinMo	-\|39.9\|46.7\|26.0
Qwen-Audio	25.1\|33.9\|41.5\|15.7
Qwen2-Audio	29.9\|35.2\|45.2\|24.4
Qwen2.5-Omni-3B	28.3\|38.1\|41.4\|26.6
Qwen2.5-Omni-7B	30.2\|37.7\|41.4\|29.4
SER
Meld	WavLM-large	0.542
MiniCPM-o	0.524
Qwen-Audio	0.557
Qwen2-Audio	0.553
Qwen2.5-Omni-3B	0.558
Qwen2.5-Omni-7B	0.570
VSC
VocalSound	CLAP	0.495
Pengi	0.604
Qwen-Audio	0.929
Qwen2-Audio	0.939
Qwen2.5-Omni-3B	0.936
Qwen2.5-Omni-7B	0.939
Music
GiantSteps Tempo	Llark-7B	0.86
Qwen2.5-Omni-3B	0.88
Qwen2.5-Omni-7B	0.88
MusicCaps	LP-MusicCaps	0.291\|0.149\|0.089\|0.061\|0.129\|0.130
Qwen2.5-Omni-3B	0.325\|0.163\|0.093\|0.057\|0.132\|0.229
Qwen2.5-Omni-7B	0.328\|0.162\|0.090\|0.055\|0.127\|0.225
Audio Reasoning
MMAU Sound \| Music \| Speech \| Avg	Gemini-Pro-V1.5	56.75\|49.40\|58.55\|54.90
Qwen2-Audio	54.95\|50.98\|42.04\|49.20
Qwen2.5-Omni-3B	70.27\|60.48\|59.16\|63.30
Qwen2.5-Omni-7B	67.87\|69.16\|59.76\|65.60
Voice Chatting
VoiceBench AlpacaEval \| CommonEval \| SD-QA \| MMSU	Ultravox-v0.4.1-LLaMA-3.1-8B	4.55\|3.90\|53.35\|47.17
MERaLiON	4.50\|3.77\|55.06\|34.95
Megrez-3B-Omni	3.50\|2.95\|25.95\|27.03
Lyra-Base	3.85\|3.50\|38.25\|49.74
MiniCPM-o	4.42\|4.15\|50.72\|54.78
Baichuan-Omni-1.5	4.50\|4.05\|43.40\|57.25
Qwen2-Audio	3.74\|3.43\|35.71\|35.72
Qwen2.5-Omni-3B	4.32\|4.00\|49.37\|50.23
Qwen2.5-Omni-7B	4.49\|3.93\|55.71\|61.32
VoiceBench OpenBookQA \| IFEval \| AdvBench \| Avg	Ultravox-v0.4.1-LLaMA-3.1-8B	65.27\|66.88\|98.46\|71.45
MERaLiON	27.23\|62.93\|94.81\|62.91
Megrez-3B-Omni	28.35\|25.71\|87.69\|46.25
Lyra-Base	72.75\|36.28\|59.62\|57.66
MiniCPM-o	78.02\|49.25\|97.69\|71.69
Baichuan-Omni-1.5	74.51\|54.54\|97.31\|71.14
Qwen2-Audio	49.45\|26.33\|96.73\|55.35
Qwen2.5-Omni-3B	74.73\|42.10\|98.85\|68.81
Qwen2.5-Omni-7B	81.10\|52.87\|99.42\|74.12

Image -> Text

Dataset	Qwen2.5-Omni-7B	Qwen2.5-Omni-3B	Other Best	Qwen2.5-VL-7B	GPT-4o-mini
MMMU_val	59.2	53.1	53.9	58.6	60.0
MMMU-Pro_overall	36.6	29.7	-	38.3	37.6
MathVista_testmini	67.9	59.4	71.9	68.2	52.5
MathVision_full	25.0	20.8	23.1	25.1	-
MMBench-V1.1-EN_test	81.8	77.8	80.5	82.6	76.0
MMVet_turbo	66.8	62.1	67.5	67.1	66.9
MMStar	64.0	55.7	64.0	63.9	54.8
MME_sum	2340	2117	2372	2347	2003
MuirBench	59.2	48.0	-	59.2	-
CRPE_relation	76.5	73.7	-	76.4	-
RealWorldQA_avg	70.3	62.6	71.9	68.5	-
MME-RealWorld_en	61.6	55.6	-	57.4	-
MM-MT-Bench	6.0	5.0	-	6.3	-
AI2D	83.2	79.5	85.8	83.9	-
TextVQA_val	84.4	79.8	83.2	84.9	-
DocVQA_test	95.2	93.3	93.5	95.7	-
ChartQA_{test Avg}	85.3	82.8	84.9	87.3	-
OCRBench_V2_en	57.8	51.7	-	56.3	-

Dataset	Qwen2.5-Omni-7B	Qwen2.5-Omni-3B	Qwen2.5-VL-7B	Grounding DINO	Gemini 1.5 Pro
Refcoco_val	90.5	88.7	90.0	90.6	73.2
Refcoco_textA	93.5	91.8	92.5	93.2	72.9
Refcoco_textB	86.6	84.0	85.4	88.2	74.6
Refcoco+_val	85.4	81.1	84.2	88.2	62.5
Refcoco+_textA	91.0	87.5	89.1	89.0	63.9
Refcoco+_textB	79.3	73.2	76.9	75.9	65.0
Refcocog+_val	87.4	85.0	87.2	86.1	75.2
Refcocog+_test	87.9	85.1	87.2	87.0	76.2
ODinW	42.4	39.2	37.3	55.0	36.7
PointGrounding	66.5	46.2	67.3	-	-

Video(without audio) -> Text

Dataset	Qwen2.5-Omni-7B	Qwen2.5-Omni-3B	Other Best	Qwen2.5-VL-7B	GPT-4o-mini
Video-MME_{w/o sub}	64.3	62.0	63.9	65.1	64.8
Video-MME_{w sub}	72.4	68.6	67.9	71.6	-
MVBench	70.3	68.7	67.2	69.6	-
EgoSchema_test	68.6	61.4	63.2	65.0	-

Zero-shot Speech Generation

Datasets	Model	Performance
Content Consistency
SEED test-zh \| test-en \| test-hard	Seed-TTS_ICL	1.11 \| 2.24 \| 7.58
Seed-TTS_RL	1.00 \| 1.94 \| 6.42
MaskGCT	2.27 \| 2.62 \| 10.27
E2_TTS	1.97 \| 2.19 \| -
F5-TTS	1.56 \| 1.83 \| 8.67
CosyVoice 2	1.45 \| 2.57 \| 6.83
CosyVoice 2-S	1.45 \| 2.38 \| 8.08
Qwen2.5-Omni-3B_ICL	1.95 \| 2.87 \| 9.92
Qwen2.5-Omni-3B_RL	1.58 \| 2.51 \| 7.86
Qwen2.5-Omni-7B_ICL	1.70 \| 2.72 \| 7.97
Qwen2.5-Omni-7B_RL	1.42 \| 2.32 \| 6.54
Speaker Similarity
SEED test-zh \| test-en \| test-hard	Seed-TTS_ICL	0.796 \| 0.762 \| 0.776
Seed-TTS_RL	0.801 \| 0.766 \| 0.782
MaskGCT	0.774 \| 0.714 \| 0.748
E2_TTS	0.730 \| 0.710 \| -
F5-TTS	0.741 \| 0.647 \| 0.713
CosyVoice 2	0.748 \| 0.652 \| 0.724
CosyVoice 2-S	0.753 \| 0.654 \| 0.732
Qwen2.5-Omni-3B_ICL	0.741 \| 0.635 \| 0.748
Qwen2.5-Omni-3B_RL	0.744 \| 0.635 \| 0.746
Qwen2.5-Omni-7B_ICL	0.752 \| 0.632 \| 0.747
Qwen2.5-Omni-7B_RL	0.754 \| 0.641 \| 0.752

Text -> Text

Dataset	Qwen2.5-Omni-7B	Qwen2.5-Omni-3B	Qwen2.5-7B	Qwen2.5-3B	Qwen2-7B	Llama3.1-8B	Gemma2-9B
MMLU-Pro	47.0	40.4	56.3	43.7	44.1	48.3	52.1
MMLU-redux	71.0	60.9	75.4	64.4	67.3	67.2	72.8
LiveBench₀₈₃₁	29.6	22.3	35.9	26.8	29.2	26.7	30.6
GPQA	30.8	34.3	36.4	30.3	34.3	32.8	32.8
MATH	71.5	63.6	75.5	65.9	52.9	51.9	44.3
GSM8K	88.7	82.6	91.6	86.7	85.7	84.5	76.7
HumanEval	78.7	70.7	84.8	74.4	79.9	72.6	68.9
MBPP	73.2	70.4	79.2	72.7	67.2	69.6	74.9
MultiPL-E	65.8	57.6	70.4	60.2	59.1	50.7	53.4
LiveCodeBench_2305-2409	24.6	16.5	28.7	19.9	23.9	8.3	18.9

Quickstart

Below, we provide simple examples to show how to use Qwen2.5-Omni with 🤗 Transformers. The codes of Qwen2.5-Omni has been in the latest Hugging face transformers and we advise you to build from source with command:

pip uninstall transformers
pip install git+https://github.com/huggingface/transformers@v4.51.3-Qwen2.5-Omni-preview
pip install accelerate

or you might encounter the following error:

KeyError: 'qwen2_5_omni'

We offer a toolkit to help you handle various types of audio and visual input more conveniently, as if you were using an API. This includes 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

# It's highly recommended to use `[decord]` feature for faster video loading.
pip install qwen-omni-utils[decord] -U

If you are not using Linux, you might not be able to install decord from PyPI. In that case, you can use pip install qwen-omni-utils -U which will fall back to using torchvision for video processing. However, you can still install decord from source to get decord used when loading video.

🤗 Transformers Usage

Here we show a code snippet to show you how to use the chat model with transformers and qwen_omni_utils:

python

import soundfile as sf

from transformers import Qwen2_5OmniForConditionalGeneration, Qwen2_5OmniProcessor
from qwen_omni_utils import process_mm_info

# default: Load the model on the available device(s)
model = Qwen2_5OmniForConditionalGeneration.from_pretrained("Qwen/Qwen2.5-Omni-3B", torch_dtype="auto", device_map="auto")

# We recommend enabling flash_attention_2 for better acceleration and memory saving.
# model = Qwen2_5OmniForConditionalGeneration.from_pretrained(
#     "Qwen/Qwen2.5-Omni-3B",
#     torch_dtype="auto",
#     device_map="auto",
#     attn_implementation="flash_attention_2",
# )

processor = Qwen2_5OmniProcessor.from_pretrained("Qwen/Qwen2.5-Omni-3B")

conversation = [
    {
        "role": "system",
        "content": [
            {"type": "text", "text": "You are Qwen, a virtual human developed by the Qwen Team, Alibaba Group, capable of perceiving auditory and visual inputs, as well as generating text and speech."}
        ],
    },
    {
        "role": "user",
        "content": [
            {"type": "video", "video": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen2.5-Omni/draw.mp4"},
        ],
    },
]

# set 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, use_audio_in_video=USE_AUDIO_IN_VIDEO)

text = processor.batch_decode(text_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)
print(text)
sf.write(
    "output.wav",
    audio.reshape(-1).detach().cpu().numpy(),
    samplerate=24000,
)

Minimum GPU memory requirements

Model	Precision	15(s) Video	30(s) Video	60(s) Video
Qwen-Omni-3B	FP32	89.10 GB	Not Recommend	Not Recommend
Qwen-Omni-3B	BF16	18.38 GB	22.43 GB	28.22 GB
Qwen-Omni-7B	FP32	93.56 GB	Not Recommend	Not Recommend
Qwen-Omni-7B	BF16	31.11 GB	41.85 GB	60.19 GB

Note: The table above presents the theoretical minimum memory requirements for inference with transformers and BF16 is test with attn_implementation="flash_attention_2"; however, in practice, the actual memory usage is typically at least 1.2 times higher. For more information, see the linked resource here.

Video URL resource usage

Video URL compatibility largely depends on the third-party library version. The details are in the table below. Change the backend by FORCE_QWENVL_VIDEO_READER=torchvision or FORCE_QWENVL_VIDEO_READER=decord if you prefer not to use the default one.

Backend	HTTP	HTTPS
torchvision >= 0.19.0	✅	✅
torchvision < 0.19.0	❌	❌
decord	✅	❌

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

# Sample messages for batch inference

# Conversation with video only
conversation1 = [
    {
        "role": "system",
        "content": [
            {"type": "text", "text": "You are Qwen, a virtual human developed by the Qwen Team, Alibaba Group, capable of perceiving auditory and visual inputs, as well as generating text and speech."}
        ],
    },
    {
        "role": "user",
        "content": [
            {"type": "video", "video": "/path/to/video.mp4"},
        ]
    }
]

# Conversation with audio only
conversation2 = [
    {
        "role": "system",
        "content": [
            {"type": "text", "text": "You are Qwen, a virtual human developed by the Qwen Team, Alibaba Group, capable of perceiving auditory and visual inputs, as well as generating text and speech."}
        ],
    },
    {
        "role": "user",
        "content": [
            {"type": "audio", "audio": "/path/to/audio.wav"},
        ]
    }
]

# Conversation with pure text
conversation3 = [
    {
        "role": "system",
        "content": [
            {"type": "text", "text": "You are Qwen, a virtual human developed by the Qwen Team, Alibaba Group, capable of perceiving auditory and visual inputs, as well as generating text and speech."}
        ],
    },
    {
        "role": "user",
        "content": "who are you?"
    }
]


# Conversation with mixed media
conversation4 = [
    {
        "role": "system",
        "content": [
            {"type": "text", "text": "You are Qwen, a virtual human developed by the Qwen Team, Alibaba Group, capable of perceiving auditory and visual inputs, as well as generating text and speech."}
        ],
    },
    {
        "role": "user",
        "content": [
            {"type": "image", "image": "/path/to/image.jpg"},
            {"type": "video", "video": "/path/to/video.mp4"},
            {"type": "audio", "audio": "/path/to/audio.wav"},
            {"type": "text", "text": "What are the elements can you see and hear in these medias?"},
        ],
    }
]

# Combine messages for batch processing
conversations = [conversation1, conversation2, conversation3, conversation4]

# set 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
text_ids = model.generate(**inputs, use_audio_in_video=USE_AUDIO_IN_VIDEO, return_audio=False)
text = processor.batch_decode(text_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)
print(text)

Usage Tips

Prompt for audio output

If users need audio output, the system prompt must be set as "You are Qwen, a virtual human developed by the Qwen Team, Alibaba Group, capable of perceiving auditory and visual inputs, as well as generating text and speech.", otherwise the audio output may not work as expected.

{
    "role": "system",
    "content": [
        {"type": "text", "text": "You are Qwen, a virtual human developed by the Qwen Team, Alibaba Group, capable of perceiving auditory and visual inputs, as well as generating text and speech."}
    ],
}

Use audio in video

In the process of multimodal interaction, the videos provided by users are often accompanied by audio (such as questions about the content in the video, or sounds generated by certain events in the video). This information is conducive to the model providing a better interactive experience. So we provide the following options for users to decide whether to use audio in video.

python

# first place, in data preprocessing
audios, images, videos = process_mm_info(conversations, use_audio_in_video=True)

python

# second place, in model processor
inputs = processor(text=text, audio=audios, images=images, videos=videos, return_tensors="pt", 
                   padding=True, use_audio_in_video=True)

python

#  third place, in model inference
text_ids, audio = model.generate(**inputs, use_audio_in_video=True)

It is worth noting that during a multi-round conversation, the use_audio_in_video parameter in these places must be set to the same, otherwise unexpected results will occur.

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 init the model. This option will save about ~2GB of GPU memory but the return_audio option for generate function will only allow to be set at False.

python

model = Qwen2_5OmniForConditionalGeneration.from_pretrained(
    "Qwen/Qwen2.5-Omni-3B",
    torch_dtype="auto",
    device_map="auto"
)
model.disable_talker()

In order to obtain a flexible experience, we recommend that users can decide whether to return audio when generate function is called. If return_audio is set to False, the model will only return text outputs to get text responses faster.

python

model = Qwen2_5OmniForConditionalGeneration.from_pretrained(
    "Qwen/Qwen2.5-Omni-3B",
    torch_dtype="auto",
    device_map="auto"
)
...
text_ids = model.generate(**inputs, return_audio=False)

Change voice type of output audio

Qwen2.5-Omni supports the ability to change the voice of the output audio. The "Qwen/Qwen2.5-Omni-3B" checkpoint support two voice types as follow:

Voice Type	Gender	Description
Chelsie	Female	A honeyed, velvety voice that carries a gentle warmth and luminous clarity.
Ethan	Male	A bright, upbeat voice with infectious energy and a warm, approachable vibe.

Users can use the speaker parameter of generate function to specify the voice type. By default, if speaker is not specified, the default voice type is Chelsie.

text_ids, audio = model.generate(**inputs, speaker="Chelsie")

text_ids, audio = model.generate(**inputs, speaker="Ethan")

Flash-Attention 2 to speed up generation

First, make sure to install the latest version of Flash Attention 2:

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 flash attention repository. FlashAttention-2 can only be used when a model is loaded in torch.float16 or torch.bfloat16.

To load and run a model using FlashAttention-2, add attn_implementation="flash_attention_2" when loading the model:

python

from transformers import Qwen2_5OmniForConditionalGeneration

model = Qwen2_5OmniForConditionalGeneration.from_pretrained(
    "Qwen/Qwen2.5-Omni-3B",
    device_map="auto",
    torch_dtype=torch.bfloat16,
    attn_implementation="flash_attention_2",
)

Citation

If you find our paper and code useful in your research, please consider giving a star :star: and citation :pencil: :)

BibTeX

@article{Qwen2.5-Omni,
  title={Qwen2.5-Omni Technical Report},
  author={Jin Xu, Zhifang Guo, Jinzheng He, Hangrui Hu, Ting He, Shuai Bai, Keqin Chen, Jialin Wang, Yang Fan, Kai Dang, Bin Zhang, Xiong Wang, Yunfei Chu, Junyang Lin},
  journal={arXiv preprint arXiv:2503.20215},
  year={2025}
}

🚀 If you find these models useful

Help me test my AI-Powered Quantum Network Monitor Assistant with quantum-ready security checks:

👉 Quantum Network Monitor

The full Open Source Code for the Quantum Network Monitor Service available at my github repos ( repos with NetworkMonitor in the name) : Source Code Quantum Network Monitor. You will also find the code I use to quantize the models if you want to do it yourself GGUFModelBuilder

💬 How to test:
Choose an AI assistant type:

TurboLLM (GPT-4.1-mini)
HugLLM (Hugginface Open-source models)
TestLLM (Experimental CPU-only)

What I’m Testing

I’m pushing the limits of small open-source models for AI network monitoring, specifically:

Function calling against live network services
How small can a model go while still handling:
- Automated Nmap security scans
- Quantum-readiness checks
- Network Monitoring tasks

🟡 TestLLM – Current experimental model (llama.cpp on 2 CPU threads on huggingface docker space):

✅ Zero-configuration setup
⏳ 30s load time (slow inference but no API costs) . No token limited as the cost is low.
🔧 Help wanted! If you’re into edge-device AI, let’s collaborate!

Other Assistants

🟢 TurboLLM – Uses gpt-4.1-mini :

**It performs very well but unfortunatly OpenAI charges per token. For this reason tokens usage is limited.
Create custom cmd processors to run .net code on Quantum Network Monitor Agents
Real-time network diagnostics and monitoring
Security Audits
Penetration testing (Nmap/Metasploit)

🔵 HugLLM – Latest Open-source models:

🌐 Runs on Hugging Face Inference API. Performs pretty well using the lastest models hosted on Novita.

💡 Example commands you could test:

"Give me info on my websites SSL certificate"
"Check if my server is using quantum safe encyption for communication"
"Run a comprehensive security audit on my server"
'"Create a cmd processor to .. (what ever you want)" Note you need to install a Quantum Network Monitor Agent to run the .net code from. This is a very flexible and powerful feature. Use with caution!

Final Word

I fund the servers used to create these model files, run the Quantum Network Monitor service, and pay for inference from Novita and OpenAI—all out of my own pocket. All the code behind the model creation and the Quantum Network Monitor project is open source. Feel free to use whatever you find helpful.

If you appreciate the work, please consider buying me a coffee ☕. Your support helps cover service costs and allows me to raise token limits for everyone.

I'm also open to job opportunities or sponsorship.

Thank you! 😊

Mungert/Qwen2.5-Omni-3B-GGUF

Qwen2.5-Omni-3B GGUF Models

Model Generation Details

Quantization beyond the IMatrix

Please provide feedback on how you find this method performs

Choosing the Right Model Format

BF16 (Brain Float 16) – Use if BF16 acceleration is available

F16 (Float 16) – More widely supported than BF16

Hybrid Precision Models (e.g., bf16_q8_0, f16_q4_K) – Best of Both Worlds

Quantized Models (Q4_K, Q6_K, Q8, etc.) – For CPU & Low-VRAM Inference

Very Low-Bit Quantization (IQ3_XS, IQ3_S, IQ3_M, Q4_K, Q4_0)

Ultra Low-Bit Quantization (IQ1_S IQ1_M IQ2_S IQ2_M IQ2_XS IQ2_XSS)

Summary Table: Model Format Selection

Qwen2.5-Omni

Overview

Introduction

Key Features

Model Architecture

Performance

Quickstart

🤗 Transformers Usage

Usage Tips

Prompt for audio output

Use audio in video

Use audio output or not

Change voice type of output audio

Flash-Attention 2 to speed up generation

Citation

🚀 If you find these models useful

What I’m Testing

Other Assistants

💡 Example commands you could test:

Final Word

Mungert/
Qwen2.5-Omni-3B-GGUF

Hybrid Precision Models (e.g., `bf16_q8_0`, `f16_q4_K`) – Best of Both Worlds