Qwen2.5-Omni: Inputs, Outputs, Checkpoints, and Licenses

Qwen2.5-Omni is an open-weight multimodal model family from the Qwen team. It can process text, images, audio, and video, then generate a text response and, when the Talker component is enabled, synthesized speech. It does not generate images or videos.

Last verified: August 23, 2026. Independence disclosure: Qwen-AI.chat is an independent informational website and demo. It is not operated, authorized, or endorsed by Alibaba, Alibaba Cloud, QwenCloud, or the Qwen team. This page separates official open checkpoints from provider-hosted API IDs and does not imply that this site’s chat interface runs this model.

Qwen2.5-Omni at a glance

Property	Verified detail
Official checkpoints	Qwen/Qwen2.5-Omni-3B and Qwen/Qwen2.5-Omni-7B
Accepted inputs	Text, images, audio, and video
Native outputs	Text and synthesized speech
Image or video generation	No
Architecture	Thinker-Talker with a speech synthesizer
Local position setting	32,768 in the official 3B and 7B checkpoint configurations
Hosted API lifecycle	The Alibaba Cloud Bailian / Chinese-mainland Model Studio table schedules qwen2.5-omni-7b for decommissioning on October 10, 2026 and names qwen3.5-omni-plus as the non-realtime replacement. QwenCloud international still listed the ID at this August 4 check.
Open-weight lifecycle	The hosted-service notice does not retire or remove the separately published 3B and 7B repositories.

The Qwen team released the 7B checkpoint on March 26, 2025 and the 3B checkpoint on April 30, 2025. The official Qwen2.5-Omni repository describes both as end-to-end multimodal models built to combine perception and speech generation.

Official checkpoints and licenses

Always identify the exact repository before describing Qwen2.5-Omni’s license. The family name alone is not enough.

Checkpoint	License	Practical meaning	Best fit
Qwen/Qwen2.5-Omni-7B	Apache License 2.0	Commercial and non-commercial use is permitted subject to the Apache 2.0 terms, notices, and applicable law.	Teams prioritizing the stronger officially reported checkpoint and a permissive license.
Qwen/Qwen2.5-Omni-3B	Qwen Research License	The published license permits non-commercial research and evaluation. Commercial use requires separate permission from the licensor.	Research, evaluation, and more resource-constrained experiments where the license is acceptable.

The “3B” and “7B” labels are part of the official checkpoint names. They should not be presented as a complete count of every parameter across all vision, audio, Thinker, Talker, and speech-synthesis components. Review the model card and license again before redistribution, fine-tuning, or production deployment.

What Qwen2.5-Omni can accept and produce

Modality	Input	Output	What that means
Text	Yes	Yes	Prompts, questions, instructions, and generated text responses.
Image	Yes	No	The model can analyze an image, but it does not render a new image.
Audio	Yes	Yes	It can interpret audio input and can synthesize a spoken response.
Video	Yes	No	It can reason over sampled video frames and the video’s audio, but it does not render a new video.

For example, an application can ask the model to summarize a short video, explain what is visible and audible, or answer a question about an image while returning text or speech. These are model capabilities, not guarantees of accuracy. Test the exact checkpoint, media type, language, prompt, and deployment stack against your own evaluation set.

How the Thinker–Talker architecture works

1. Perception: dedicated encoders process visual and audio information alongside text.
2. Thinker: the main multimodal component reasons over the combined inputs and generates text plus representations used by the speech path.
3. Talker: an autoregressive component predicts streaming speech tokens conditioned on the Thinker’s output.
4. Speech synthesis: the Token2Wav component converts the speech representation into an audible waveform.

The architecture also uses time-aligned multimodal rotary position embedding (TMRoPE) to align audio and video along a shared timeline. This design is documented in the Qwen2.5-Omni Technical Report.

Context length: repository versus hosted route

The official 7B configuration and 3B configuration each set max_position_embeddings to 32,768. Media consumes processed multimodal tokens, so this is not a promise of 32,768 text tokens plus unrestricted audio or video.

Alibaba Cloud’s general Model Studio matrix lists a 1M context for the hosted qwen2.5-omni-7b ID, while QwenCloud’s current international vision matrix lists a 32K context and 2K maximum output. The official 3B and 7B checkpoint configurations also use 32,768 positions. Treat 32K as the conservative documented limit unless the exact regional console and endpoint explicitly confirm a higher hosted-service limit. Never apply a hosted-service limit to a local checkpoint configuration.

The Chinese-mainland/Bailian hosted ID is scheduled to decommission on October 10, 2026. QwenCloud international still listed the same string and did not list it on its current deprecation page at this check. Confirm the base URL, key, deployment scope, account notice, and model card together.

Local checkpoint names and hosted model IDs are not interchangeable

Name	Where it belongs	Verified status on August 23, 2026
Qwen/Qwen2.5-Omni-7B	Official Hugging Face open-weight checkpoint	Repository remains published under Apache 2.0; it is not retired by a hosted API notice
Qwen/Qwen2.5-Omni-3B	Official Hugging Face open-weight checkpoint	Repository remains published under the Qwen Research License
qwen2.5-omni-7b	Provider-hosted API ID	Bailian / Chinese-mainland Model Studio: scheduled for October 10, 2026 decommissioning, with qwen3.5-omni-plus named as replacement. QwenCloud international still listed the ID; verify that deployment separately.
qwen-omni-turbo	Separate provider-hosted model ID	Not an alias for the open Qwen2.5-Omni repository; check its own route-specific lifecycle

For a new hosted project, compare the current Omni route in the provider and region you will actually use. See the Qwen3-Omni model guide and the Qwen model selector.

Run Qwen2.5-Omni locally with Transformers

This minimal example uses the exact 7B checkpoint and returns text from a video containing both frames and audio. Install FFmpeg first. The official repository currently pins Transformers 4.52.3, so verify that version again before building a production image.

python -m pip install "transformers==4.52.3" accelerate
python -m pip install -U "qwen-omni-utils[decord]"

from transformers import (
    Qwen2_5OmniForConditionalGeneration,
    Qwen2_5OmniProcessor,
)
from qwen_omni_utils import process_mm_info

model_id = "Qwen/Qwen2.5-Omni-7B"

model = Qwen2_5OmniForConditionalGeneration.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)
processor = Qwen2_5OmniProcessor.from_pretrained(model_id)

conversation = [{
    "role": "user",
    "content": [
        {"type": "video", "video": "/absolute/path/input.mp4"},
        {"type": "text", "text": "Summarize what is visible and audible."},
    ],
}]

prompt = processor.apply_chat_template(
    conversation,
    add_generation_prompt=True,
    tokenize=False,
)
audios, images, videos = process_mm_info(
    conversation,
    use_audio_in_video=True,
)
inputs = processor(
    text=prompt,
    audio=audios,
    images=images,
    videos=videos,
    return_tensors="pt",
    padding=True,
    use_audio_in_video=True,
)
inputs = inputs.to(model.device).to(model.dtype)

text_ids = model.generate(
    **inputs,
    use_audio_in_video=True,
    return_audio=False,
)
reply = processor.batch_decode(
    text_ids,
    skip_special_tokens=True,
    clean_up_tokenization_spaces=False,
)
print(reply[0])

For speech output, follow the official model card’s documented system prompt and call generate() with audio enabled. The 7B model card documents the speakers Chelsie and Ethan, with Chelsie as the default. If speech is not needed, model.disable_talker() reduces memory use and forces text-only generation.

GPU memory requirements

The following BF16 values are the Qwen team’s theoretical minimum GPU memory for its documented video tests. The repository warns that real usage is typically at least 1.2 times higher. Input resolution, media duration, batch size, attention implementation, framework overhead, and whether speech generation is enabled can all change the result.

Checkpoint	15-second video	30-second video	60-second video
Qwen2.5-Omni-3B, BF16	18.38 GB	22.43 GB	28.22 GB
Qwen2.5-Omni-7B, BF16	31.11 GB	41.85 GB	60.19 GB

Do not promise that the 7B BF16 model fits a 24 GB GPU for multimodal video inference. The Qwen team also publishes official 4-bit AWQ and GPTQ-Int4 variants of the 7B checkpoint that reduce VRAM substantially in its tests, but quantization, CPU offloading, and text-only mode can affect quality and speed. Benchmark your actual workload before choosing hardware.

Deployment options and important limits

• Transformers: the most direct documented path for local text and speech generation.
• Official Docker image: useful when you want the dependency versions maintained by the Qwen repository.
• vLLM: the repository documents audio output for offline inference. Its documented vllm serve path is Thinker-only and returns text, so verify the current implementation before promising a speech API.
• MNN: an official edge-deployment path with its own supported platforms and benchmarks.
• Alibaba Cloud Model Studio: the hosted Qwen2.5 IDs are legacy. Check the current Qwen API guide and official provider documentation before integrating.

The current Model Studio model table marks thinking mode, function calling, built-in tools, and structured output as unsupported for hosted qwen2.5-omni-7b. Do not market those features as native Qwen2.5-Omni capabilities. An application can add external orchestration, validation, or retrieval, but those are properties of the application stack—not proof of native model support.

Where Qwen2.5-Omni can be useful

• Audio-visual question answering over short clips.
• Video summaries that consider both frames and sound.
• Spoken responses grounded in an image, recording, or video.
• Prototyping multimodal assistants in research or controlled product tests.
• Comparing open-weight multimodal architectures and deployment methods.

These are candidate use cases, not endorsements for autonomous decisions. Model outputs can be incomplete or wrong. Add human review and task-specific testing for medical, legal, financial, safety, security, accessibility, or other high-impact uses. Never assume that speech fluency proves factual accuracy.

Should you choose 3B, 7B, or a newer Omni model?

• Choose the 3B checkpoint for non-commercial research or evaluation where lower resource requirements matter and the Qwen Research License fits the use.
• Choose the 7B checkpoint when you want the primary Qwen2.5-Omni open checkpoint, can support its higher memory requirements, and need Apache 2.0 terms.
• Migrate the Chinese-mainland/Bailian hosted ID before October 10, 2026; its lifecycle table names qwen3.5-omni-plus as the non-realtime replacement.
• Verify QwenCloud international separately because it still listed qwen2.5-omni-7b and did not include it on its deprecation page at the August 4 check.

Run representative tests for languages, media lengths, latency, output mode, factuality requirements, and hardware. This update is a documentation review; no authenticated hosted request or local benchmark was completed.

Official Qwen2.5-Omni sources

• Official Qwen2.5-Omni release post
• Official GitHub repository and deployment notes
• Official Qwen2.5-Omni-7B model card
• Official Qwen2.5-Omni-3B model card
• Qwen2.5-Omni technical report
• Alibaba Cloud Bailian / Model Studio lifecycle table
• QwenCloud international vision-model matrix
• QwenCloud international deprecation log
• Alibaba Cloud Model Studio model matrix

Editorial verification: Facts and outbound links on this page were checked against the first-party sources above on August 23, 2026. This was a documentation review, not a live inference or billing-console test.

Frequently asked questions