TaoLiveAIGC/TLive-Omni-4B

TLive-Omni

📋 Overview

TLive-Omni is an omni-modal understanding model for e-commerce live-stream, mapping image, video, audio, and text into a unified text-output interface. Built on a Qwen3.5 backbone with a grafted AuT audio encoder, it supports up to 256K tokens of context, trained via a three-stage SFT recipe followed by Faithful-RFT reinforcement fine-tuning.

✨ Highlights

• Timestamped Per-vGrid layout — Audio and video tokens are organized into timestamped grid with explicit boundaries, keeping audio segments adjacent to their corresponding visual content for fine-grained temporal alignment over long streams.
• Three-stage SFT recipe — Progressive training from audio-language alignment to full multimodal SFT, developing live-commerce understanding from omni-modal perception to instruction-following responses.
• Faithful-RFT — A reinforcement fine-tuning stage for faithful and real-time live-stream demands, suppressing explicit reasoning traces and directly optimizing answer quality for live-commerce tasks.
• Rich atomic capabilities — A scenario-oriented taxonomy covering speech recognition, speaker analysis, product visual grounding, text recognition, temporal grounding, video dense caption, and omni-modal QA, etc, supported by a compact data production engine.
• Strong live-commerce performance with competitive generalization — 4B and 9B variants demonstrate strong results across live-commerce audio, image, and video tasks, together with excellent generalization on general benchmarks.

🏗️ Architecture

TLive-Omni is built on a Qwen3.5 backbone and extends it with a audio encoder through a lightweight MLP aligner, forming a unified text-output omni-modal understanding model. For video inputs with audio, each temporal grid is organized into a timestamped grid that interleaves video and audio token blocks, keeping audio segments adjacent to their corresponding visual content. The model supports up to 256K tokens of context at inference.

📊 Benchmark Results

We evaluate TLive-Omni-4B and TLive-Omni-9B on both live-commerce tasks and general benchmarks. Dash (-) denotes an unreported result or undisclosed parameter count. The Best results among the compared open-source models are marked in bold, while the second-best results are in underlined.

Live-Commerce Evaluation

Click to expand

Task	Metric	TLive-Omni 4B	TLive-Omni 9B	Gemini 2.5 Flash	Gemini 2.5 Pro	Gemini 3 Flash	Gemini 3 Pro	Gemini 3.5 Flash	Qwen3.5-Omni Flash	OmniVinci 9B	Nemotron 3 Nano Omni 30B-A3B	Ming-Lite-Omni v1.5 20B-A3B	MiniCPM-o 2.6 8B	MiniCPM-o 4.5 9B	Qwen2.5-Omni 7B	Qwen3-Omni 30B-A3B
Audio
Live-Commerce ASR	CER ↓	6.66	6.46	16.30	11.48	15.18	12.09	13.09	6.81	—	12.10	10.06	13.88	10.70	7.86	6.75
Speaker-Attributed ASR	cpWER ↓	12.88	12.27	17.14	12.17	19.04	11.67	11.99	13.23	—	17.65	—	—	18.89	—	27.84
Audio Description	Acc. ↑	76.12	75.96	65.21	81.10	68.27	85.07	79.97	62.82	39.90	33.01	45.99	49.84	47.59	47.92	61.06
Audio Description	Hal. ↓	20.97	21.00	26.19	14.16	26.17	10.92	14.36	27.81	47.36	39.77	44.05	41.76	52.41	36.92	30.22
Audio QA	Acc. ↑	72.60	76.28	76.28	82.85	74.68	88.62	87.99	78.04	66.51	64.90	40.54	39.74	42.47	61.38	76.76
Image
Visual Grounding	Live AP ↑	82.85	82.33	61.08	51.98	80.38	73.80	84.15	79.96	34.86	73.08	52.46	3.82	23.90	75.61	79.22
Visual Grounding	Prod AP ↑	91.45	89.96	28.81	32.63	65.67	58.83	74.89	60.44	8.93	48.62	40.73	1.77	53.63	22.85	68.88
Text Understanding	Loc. F1 ↑	86.99	87.59	20.52	31.60	61.11	68.60	64.44	74.07	50.25	52.91	13.27	5.74	5.43	42.64	30.46
Text Understanding	Rec. NED ↓	4.72	4.24	43.28	27.82	16.25	9.72	16.64	12.48	32.77	29.42	59.16	77.58	71.65	37.79	14.83
Text Understanding	Cls. Acc. ↑	79.06	79.85	51.21	61.86	69.11	76.86	69.76	53.25	57.29	37.86	32.94	15.92	11.62	51.25	69.46
Video
Temporal Grounding	mIoU ↑	77.63	81.49	76.50	76.22	77.43	77.90	77.90	62.10	13.10	23.39	14.34	14.56	43.20	30.83	39.22
Dense Caption	Acc. ↑	69.23	74.63	54.60	41.95	32.21	37.80	33.80	32.94	18.59	17.96	13.81	10.53	21.06	16.51	21.44
Dense Caption	Hal. ↓	9.57	8.76	10.97	16.88	20.76	20.99	17.30	20.91	27.13	16.62	39.33	26.93	28.61	36.44	25.82
Video QA	Acc. ↑	92.31	93.23	88.21	92.62	89.64	84.36	86.90	87.28	72.51	82.56	64.51	60.30	84.62	75.48	81.62
Shot Understanding	Layout ↑	78.40	77.00	80.00	85.20	76.80	80.40	83.40	84.40	73.60	79.20	74.60	66.60	78.20	74.40	82.20
Shot Understanding	Shot Size ↑	51.20	51.00	46.80	50.80	45.70	43.40	44.20	48.90	52.70	34.00	41.70	38.30	42.80	38.10	37.40
Shot Understanding	Camera ↑	80.90	82.00	84.20	76.00	78.50	75.70	75.50	85.50	68.10	80.20	72.80	68.30	79.20	81.20	76.10
Shot Understanding	Content ↑	69.80	71.00	68.60	70.40	71.60	74.80	70.20	66.40	49.20	58.60	51.60	49.00	66.60	68.00	63.60

General Benchmark: Image Understanding

Click to expand

Model	Params	MMMU	MathVista	DynaMath	VLMsAreBlind	MMBench(EN-DEV-v1.1)	RealWorldQA	MMStar	SimpleVQA	Hallusion	AI2D	OCRBench	CC-OCR	CharXiv(RQ)	RefCOCO	ERQA	EmbSpatial
Open-source VLM models
MiMo-VL-SFT	7B	64.6	81.8	46.9	78.0	84.5	—	—	—	—	83.2	87.6	—	54.4	85.7	—	—
SAIL-VL2	8B	55.4	76.4	17.8	—	—	76.3	70.7	—	55.1	87.7	91.3	—	—	74.0	—	—
Valley2.5	8B	62.1	74.4	32.7	—	85.5	70.5	67.3	—	56.3	84.4	87.0	—	—	—	—	—
LLaVA-OneVision-2	8B	—	—	—	—	85.7	69.7	64.8	—	—	84.3	78.2	—	—	—	43.3	78.1
InternVL3.5	4B	66.6	77.1	35.7	—	80.3	66.3	65.0	—	44.8	82.6	82.2	—	39.6	89.4	38.5	—
InternVL3.5	8B	73.4	78.4	37.7	—	79.5	67.5	69.3	—	54.5	84.0	84.0	—	44.4	89.7	41.0	73.2
Qwen3-VL	4B	67.4	73.7	65.3	71.9	83.9	70.9	69.8	48.0	57.6	84.1	88.1	76.2	39.7	89.0	41.3	79.6
Qwen3-VL	8B	69.6	77.2	67.7	74.0	84.5	71.5	70.9	50.2	61.1	85.7	89.6	79.9	46.4	89.1	45.8	78.5
Qwen3.5	4B	72.1	81.0	69.6	62.3	86.3	72.5	74.8	44.6	76.9	87.1	85.9	71.1	62.9	87.6	46.8	76.6
Qwen3.5	9B	74.2	82.2	74.6	71.8	87.7	72.9	76.3	48.9	76.0	88.0	88.5	73.4	67.5	90.0	47.3	78.7
Open-source Omni models
InteractiveOmni	4B	61.1	61.7	—	—	78.9	—	62.6	—	52.2	83.8	80.0	—	—	—	—	—
InteractiveOmni	8B	66.9	68.0	—	—	81.4	—	66.8	—	61.3	84.3	83.7	—	—	—	—	—
VITA-1.5	7B	52.1	66.2	—	—	76.7	—	59.9	—	44.9	79.3	73.2	—	—	—	—	—
Valley3	8B	69.3	—	—	—	—	—	—	—	55.9	—	—	—	—	—	—	—
OmniVinci	9B	49.7	63.5	—	—	—	67.5	—	—	—	—	—	—	—	—	—	—
Nemotron 3 Nano Omni	30B-A3B	55.2	71.9	—	—	—	—	—	—	—	88.5	88.3	—	49.1	80.6	—	—
Ming-Lite-Omni v1.5	20B-A3B	54.3	72.0	—	—	—	—	65.1	—	54.6	84.9	88.9	—	—	87.8	—	—
MiniCPM-o 2.6	8B	50.4	71.9	—	—	80.5	—	64.0	—	51.9	85.8	89.7	—	—	—	—	—
MiniCPM-o 4.5	9B	67.6	—	—	—	87.6	—	73.1	—	63.2	87.6	87.6	—	—	—	—	—
Qwen2.5-Omni	7B	59.2	67.9	—	—	81.8	70.3	64.0	—	—	83.2	—	—	—	87.7	—	—
Qwen3-Omni	30B-A3B	69.1	75.9	—	—	—	—	68.5	—	59.7	85.2	86.0	—	61.1	—	—	—
Ours
TLive-Omni	4B	70.9	79.9	72.5	71.8	87.0	77.7	73.9	47.6	77.7	86.6	86.6	80.5	61.3	87.4	42.3	79.3
TLive-Omni	9B	73.4	81.9	73.3	75.5	88.9	76.6	75.1	50.0	76.0	88.6	90.3	81.3	63.1	90.0	48.0	80.4

General Benchmark: Video Understanding

Click to expand

Model	Params	MVBench	MLVU	Video-MME	LongVideoBench	LVBench	MMVU	VideoMMMU	Charades-TL	ActivityNet-TL	QVHighlights-TL
Open-source VLM models
MiMo-VL-SFT	7B	—	—	66.9	—	—	—	53.1	39.6	35.5	41.5
SAIL-VL2	8B	—	—	62.7	58.3	—	—	—	—	—	—
LLaVA-OneVision-2	8B	66.2	76.6	71.9	66.9	55.5	56.2	—	53.5	53.8	66.4
LLaVA-Video	7B	58.6	70.8	63.3	58.2	44.2	47.1	36.1	15.2	14.6	10.4
InternVL3.5	4B	71.2	70.4	65.4	60.8	43.2	47.6	57.6	16.0	14.9	17.7
InternVL3.5	8B	72.1	70.2	66.0	62.1	46.7	60.2	—	27.8	31.3	31.3
MiniCPM-V 4.5	8B	—	75.1	67.9	63.9	50.4	58.9	57.1	31.9	32.3	46.1
LongVU	7B	66.9	65.4	60.6	—	—	—	—	—	—	—
LongVILA	7B	67.1	—	60.1	57.1	—	—	—	—	—	—
Mage-VL	4B	65.1	68.7	64.0	61.3	41.8	—	—	50.7	45.4	57.4
Molmo2	4B	75.1	63.0	69.6	68.0	53.9	51.2	50.7	33.3	39.8	58.7
Molmo2	8B	75.9	60.2	69.9	67.5	52.8	—	—	—	—	—
NVILA	8B	68.1	70.1	64.2	57.7	—	—	—	—	—	—
Kangaroo	8B	61.1	61.0	56.0	54.8	39.4	—	—	—	—	—
Video-XL2	8B	—	74.8	66.6	61.0	48.4	50.0	39.9	38.9	30.0	46.2
VideoChat3	4B	—	—	70.1	—	56.7	56.4	57.4	56.1	54.6	67.0
VideoLLaMA 3	7B	69.7	73.0	66.2	59.8	45.3	44.1	34.6	39.8	29.8	36.9
Qwen3-VL	4B	68.9	75.3	69.3	—	56.2	50.5	56.2	46.4	48.2	58.7
Qwen3-VL	8B	68.7	78.1	71.4	—	58.0	58.7	65.3	48.3	46.8	59.4
Qwen3.5	4B	66.6	75.1	71.6	65.1	55.3	57.8	69.8	48.7	51.6	55.0
Qwen3.5	9B	75.7	79.7	66.9	67.9	60.9	63.7	70.3	52.0	54.0	57.2
Open-source Omni models
InteractiveOmni	4B	—	68.0	63.3	57.0	—	—	—	—	—	—
InteractiveOmni	8B	—	71.6	66.0	59.1	—	—	—	—	—	—
VITA-1.5	7B	55.4	—	56.1	—	—	—	—	—	—	—
Valley3	8B	—	55.6	—	—	—	—	61.2	—	—	—
OmniVinci	9B	70.6	—	68.2	61.3	—	—	—	—	—	—
Nemotron 3 Nano Omni	30B-A3B	—	—	70.8	—	—	—	—	—	—	—
Ming-Lite-Omni v1.5	20B-A3B	69.4	—	67.1	59.5	—	—	—	—	—	—
MiniCPM-o 2.6	8B	—	—	63.9	—	—	—	—	—	—	—
MiniCPM-o 4.5	9B	—	76.5	70.4	66.0	—	—	—	—	—	—
Qwen2.5-Omni	7B	70.3	—	64.3	—	—	—	—	—	—	—
Qwen3-Omni	30B-A3B	—	75.2	70.5	—	—	—	—	—	—	—
Ours
TLive-Omni	4B	69.0	76.1	71.3	66.1	57.1	59.9	73.9	57.0	58.2	69.2
TLive-Omni	9B	72.5	80.9	75.6	69.9	60.8	67.1	72.8	56.3	55.4	64.1

General Benchmark: Omni Understanding

Click to expand

Model	Params	AVUT	WorldSense	VideoHolmes	DailyOmni	OmniVideoBench	FutureOmni
Open-source Omni models
video-SALMONN 2+	3B	66.2	48.3	42.2	67.7	—	—
video-SALMONN 2+	7B	69.5	50.9	46.9	71.8	—	—
OmniVinci	9B	—	48.2	—	66.5	36.7	52.8
Nemotron 3 Nano Omni	30B-A3B	—	55.2	—	74.5	—	—
MiniCPM-o 4.5	9B	78.6	55.7	64.3	80.2	41.1	56.1
Qwen2.5-Omni	7B	—	45.4	—	62.4	36.5	48.9
Qwen3-Omni	30B-A3B	74.2	54.0	50.4	71.9	43.8	53.4
Ours
TLive-Omni	4B	78.6	54.0	57.5	78.6	41.6	57.2
TLive-Omni	9B	80.0	56.0	59.3	80.5	43.2	58.5

⚙️ Installation

This release targets Python 3.10 on Linux x86_64 with CUDA 12.8 and PyTorch 2.10.0.

conda create -n tlive python=3.10 -y
conda activate tlive
pip install -r https://raw.githubusercontent.com/TaoLiveAIGC/TLive-Omni/v1.0.0-rc1/environments/requirements.txt

The remote environments/requirements.txt includes custom wheels for the supported environment and model. If any wheel does not match your hardware, CUDA version, or Python version, replace it with a compatible build for your setup.

🚀 Quick Start

Transformers inference

import torch
from transformers import AutoModelForCausalLM, AutoProcessor

model_id = "TaoLiveAIGC/TLive-Omni-4B"

processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    trust_remote_code=True,
    dtype=torch.bfloat16,
    device_map="auto",
    attn_implementation="flash_attention_2",
).eval()


def generate(messages, *, use_audio_in_video=False, videos_kwargs=None, generation_kwargs=None):
    inputs = processor.apply_chat_template(
        messages,
        add_generation_prompt=True,
        tokenize=True,
        return_dict=True,
        return_tensors="pt",
        enable_thinking=False,
        use_audio_in_video=use_audio_in_video,
        videos_kwargs=videos_kwargs or {},
    )
    prompt_length = inputs["input_ids"].shape[-1]
    inputs = inputs.to(model.device)

    generation_kwargs = generation_kwargs or {}
    with torch.inference_mode():
        generated_ids = model.generate(
            **inputs,
            do_sample=False,
            max_new_tokens=1024,
            **generation_kwargs,
        )

    answer_ids = generated_ids[:, prompt_length:]
    answer = processor.batch_decode(
        answer_ids,
        skip_special_tokens=True,
        clean_up_tokenization_spaces=False,
    )[0]
    return answer.strip()

Replace messages with one of the examples below for text, image, audio, or video inputs.

Text

messages = [{
    "role": "user",
    "content": [{"type": "text", "text": "Briefly explain why multimodal context can improve an answer."}],
}]
print(generate(messages))

Image

messages = [{
    "role": "user",
    "content": [
        {"type": "image", "path": "https://raw.githubusercontent.com/TaoLiveAIGC/TLive-Omni/main/data/image.jpg"},
        {"type": "text", "text": "Describe this image."},
    ],
}]
print(generate(messages))

Audio

messages = [{
    "role": "user",
    "content": [
        {"type": "audio", "path": "https://raw.githubusercontent.com/TaoLiveAIGC/TLive-Omni/main/data/audio.mp3"},
        {"type": "text", "text": "Transcribe and summarize this audio."},
    ],
}]
print(generate(messages))

Video with audio

messages = [{
    "role": "user",
    "content": [
        {"type": "video", "path": "https://raw.githubusercontent.com/TaoLiveAIGC/TLive-Omni/main/data/vocal_video.mp4"},
        {"type": "text", "text": "Describe the video, including relevant speech and sounds."},
    ],
}]
print(generate(messages, use_audio_in_video=True, videos_kwargs={"fps": 1.0}))

Video without audio

messages = [{
    "role": "user",
    "content": [
        {"type": "video", "path": "https://raw.githubusercontent.com/TaoLiveAIGC/TLive-Omni/main/data/silence_video.mp4"},
        {"type": "text", "text": "Describe the visual events in this video."},
    ],
}]
print(generate(messages, use_audio_in_video=False, videos_kwargs={"fps": 1.0}))

For temporal localization outputs, we recommend the MM:SS - MM:SS interval format, for example 01:23 - 01:35. For videos, set use_audio_in_video=True when the audio track should be used, and False for visual-only inference.

⚡ vLLM

Installation

First install the pre-built wheel (Python 3.10 + CUDA 12.8 + Linux x86_64), built and tested on NVIDIA H20 GPUs (Hopper, sm_90):

pip install https://github.com/TaoLiveAIGC/TLive-Omni/releases/download/v1.0.0-rc1/vllm-0.19.0+cu128-cp310-cp310-linux_x86_64.whl

If your GPU, driver, or CUDA setup is not compatible with this wheel, build vLLM from source using the customized code in the vllm/ directory of the GitHub release.

Inference

from transformers import AutoProcessor
from vllm import LLM, SamplingParams
from vllm.model_executor.models.tlive_omni_processing import process_audio_info

model_id = "TaoLiveAIGC/TLive-Omni-4B"
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)


def build_prompt(messages):
    return processor.apply_chat_template(
        messages,
        add_generation_prompt=True,
        tokenize=False,
        enable_thinking=False,
    )


def generate(inputs, *, limit_mm_per_prompt=None, vllm_kwargs=None, sampling_kwargs=None):
    vllm_kwargs = vllm_kwargs or {}
    sampling_kwargs = sampling_kwargs or {}
    llm = LLM(
        model=model_id,
        trust_remote_code=True,
        dtype="bfloat16",
        max_model_len=32768,
        tensor_parallel_size=1,
        gpu_memory_utilization=0.9,
        max_num_seqs=4,
        max_num_batched_tokens=32768,
        seed=42,
        limit_mm_per_prompt=limit_mm_per_prompt,
        **vllm_kwargs,
    )
    outputs = llm.generate(
        inputs,
        sampling_params=SamplingParams(
            temperature=0.0,
            max_tokens=1024,
            **sampling_kwargs,
        ),
    )
    return outputs[0].outputs[0].text.strip()

Replace messages with one of the examples below for text, image, audio, or video inputs.

Text

messages = [{
    "role": "user",
    "content": [{"type": "text", "text": "Briefly explain why multimodal context can improve an answer."}],
}]
inputs = {"prompt": build_prompt(messages)}
print(generate(inputs))

Image

image_path = "https://raw.githubusercontent.com/TaoLiveAIGC/TLive-Omni/main/data/image.jpg"
messages = [{
    "role": "user",
    "content": [
        {"type": "image", "path": image_path},
        {"type": "text", "text": "Describe this image."},
    ],
}]
inputs = {
    "prompt": build_prompt(messages),
    "multi_modal_data": {"image": [image_path]},
}
print(generate(inputs, limit_mm_per_prompt={"image": 1}))

Audio

audio_path = "https://raw.githubusercontent.com/TaoLiveAIGC/TLive-Omni/main/data/audio.mp3"
messages = [{
    "role": "user",
    "content": [
        {"type": "audio", "audio": audio_path},
        {"type": "text", "text": "Transcribe and summarize this audio."},
    ],
}]
inputs = {
    "prompt": build_prompt(messages),
    "multi_modal_data": {"audio": process_audio_info(messages, use_audio_in_video=False)},
}
print(generate(inputs, limit_mm_per_prompt={"audio": 1}))

Video with audio

video_path = "https://raw.githubusercontent.com/TaoLiveAIGC/TLive-Omni/main/data/vocal_video.mp4"
messages = [{
    "role": "user",
    "content": [
        {"type": "video", "video": video_path},
        {"type": "text", "text": "Describe the video, including relevant speech and sounds."},
    ],
}]
inputs = {
    "prompt": build_prompt(messages),
    "multi_modal_data": {
        "video": [video_path],
        "audio": process_audio_info(messages, use_audio_in_video=True),
    },
    "mm_processor_kwargs": {"videos_kwargs": {"fps": 1.0, "use_audio_in_video": True, "return_metadata": True}},
}
print(generate(inputs, limit_mm_per_prompt={"video": 1, "audio": 1}))

Video without audio

video_path = "https://raw.githubusercontent.com/TaoLiveAIGC/TLive-Omni/main/data/silence_video.mp4"
messages = [{
    "role": "user",
    "content": [
        {"type": "video", "video": video_path},
        {"type": "text", "text": "Describe the visual events in this video."},
    ],
}]
inputs = {
    "prompt": build_prompt(messages),
    "multi_modal_data": {"video": [video_path]},
    "mm_processor_kwargs": {"videos_kwargs": {"fps": 1.0, "use_audio_in_video": False, "return_metadata": True}},
}
print(generate(inputs, limit_mm_per_prompt={"video": 1}))

📖 Citation

If you find our work helpful, please consider citing our paper:

@article{tliveomni,
  title   = {TLive-Omni: An Omni-Modal Understanding Model for E-Commerce Live Streaming},
  author  = {TLive-Omni Team},
  journal = {arXiv preprint arXiv:XXXX.XXXXX},
  year    = {2026}
}

📄 License

This project is released under the Apache License 2.0.