Skip to main content
Back

Qwen Unveils Qwen2.5-Omni: Voice & Video Chat Now Local

TRY NOW

March 27, 2025 6 min read

Hey,

Welcome to AlphaSignal – the most read newsletter by AI developers. 


We bring you the top 1% of news, papers, models, and repos, all summarized to keep you updated on the latest in AI.

IN TODAY'S SIGNAL

Read time: 6 min 02 sec

🎖️ Top News

⚡️ Trending Signals

📌  Assembly AI

💻  Top Lectures

🧠  Python Tip

  • Customize ChatGPT responses efficiently using system messages for domain-specific tasks without retraining the model.

If you're enjoying AlphaSignal please forward this email to a colleague. 

It helps us keep this content free.

TOP NEWS

Multimodal Model

Qwen releases an open-source model which outperforms Gemini 1.5 Pro on multimodal reasoning

⇧ 4,485 Likes

What's New

Alibaba has open-sourced Qwen2.5-Omni, a model that processes and generates text, images, audio, and video. It achieves state-of-the-art performance on multimodal benchmarks and operates in real time. The Thinker-Talker architecture separates reasoning from speech synthesis for more structured outputs.



Multimodal Processing and Output
Qwen2.5-Omni handles text, images, audio, and video as input and generates text and speech.

  • Processes mixed-modal queries in a single prompt.

  • Follows speech-based instructions at a level comparable to text inputs.

  • Uses TMRoPE to synchronize timestamps across different modalities.


Architecture

Uses Thinker-Talker for end-to-end multimodal processing. A dual-track autoregressive design separates reasoning and speech synthesis for structured outputs.

The model splits processing and speech output into two components.

  • Thinker functions as a language model handling reasoning and text generation.

  • Talker generates speech based on text or direct audio instructions.

  • Components share context and run as one model during training and inference.

  • System supports end-to-end processing without separate pipelines.

Real-Time and Streaming Capabilities

The model supports real-time processing and low-latency inference.
  • Block-wise processing enables streaming inputs for continuous response generation.

  • Efficient memory management allows handling long audio and video sequences.

Benchmarks and Performance
Qwen2.5-Omni achieves strong results across multimodal reasoning and language tasks.

  • OmniBench: 56.1 – general-purpose multimodal reasoning; Gemini-1.5-Pro scores 42.9

  • MMAU: 65.6 – audio understanding; Qwen2-Audio scores 49.2

  • MVBench: 70.3 – video understanding; Qwen2.5-VL scores 69.6

  • Seed-tts-eval: 93.5 – naturalness of speech generation; human baseline is 93.2

  • NMOS+: 4.51 – subjective speech quality (Mean Opinion Score); same as human score


Availability and Access

You can access everything and run it without cloud services.

  • Released under Apache 2.0 with commercial-use permission.

  • Available on GitHub, Hugging Face, and ModelScope.

  • Includes inference code and technical documentation.

  • Supports local deployment without API dependencies.


Community Feedback

Niels Rogge
"Not only did they release the 7B model behind qwen chat 's voice mode with an Apache 2.0 license, integrate it into Transformers from day 1, they also have an entire technical report where they describe everything"

Dr. Daniel Bender
"How do you run it? Along the instructions on the Hugging face page? The memory requirements to process video are huge: BF16 31.11 GB (15 sec) 41.85 GB (40 sec) 60.19 GB (60 sec)"

Leonardo Silva
"When will you make a dedicated app available for Qwen AI ??"

Try now

TRENDING SIGNALS

AI Tools

OpenAI adds Anthropic's MCP to Agents SDK, streamlining third-party tool integration into AI workflows

⇧ 12,573 Likes

AI Tools

Anthropic updates MCP spec with OAuth 2.1, streamable HTTP transport, and enhances tool annotations for better tool integration

⇧ 1,538 Likes

Image Generation

Ideogram announces v3.0, offers photorealism, advanced text rendering, and style references for precise design

⇧ 1,264 Likes

AI Tools

Google DeepMind rolls out new Gemini function calling guide with full Python, JavaScript, and REST examples

⇧ 426 Likes

Research Tools

Ai2 introduces Paper Finder, an LLM tool for deep litrature discovery and citation tracing through iterative querying

⇧ 863 Like

The Most Accurate Speech AI for Conversation Intelligence

AssemblyAI sets the standard for conversation intelligence, delivering unrivaled transcription, speaker diarization, and advanced speech understanding.

Reduce transcript errors, generate actionable insights, and power smarter downstream processes.

Build with enterprise-grade Speech AI that scales seamlessly and stays ahead of the competition.

Try for Free ↗️

TOP LECTURES

Coding Assistance

Vibe Coding 101 with Replit

⇧ 3,483 Likes

In this course, you'll learn how to build and deploy applications with AI agents using Replit's cloud environment. Discover effective coding strategies like modular problem-solving, precise prompts, and structured debugging. You'll also create prototypes, use agents for design, and develop a national park ranking app, refining your workflow.

AI Development

Future of Jobs Report 2025 by World Economic Forum

⇧ 10,482 Likes

In this report, you'll learn how AI is reshaping the labor market and the skills needed for 2030. The focus is on upskilling in AI and big data, fostering creativity, and staying adaptable. You’ll explore the top skills for the future and discover how to incorporate AI tools in your field.

AI Model Inference

How to run DeepSeek-V3-0324 locally

⇧ 980 Likes

Learn how to run DeepSeek-V3-0324, a model rivaling GPT-4.5, on your local device. This session covers the use of Unsloth AI's Dynamic GGUFs, optimizing performance, and the differences between standard and dynamic quantization. Gain hands-on knowledge of running inference efficiently on GPU or CPU setups.

PYTHON TIP

Streamline Model Fine-Tuning with Custom System Messages in the ChatGPT SDK

Leverage the openai.ChatCompletion.create method with custom system messages for efficient model fine-tuning. By passing specialized prompts or instructions in the system field, you can adapt ChatGPT for specific tasks (e.g., technical explanations, research summaries) without retraining from scratch.


How This Works

By setting the system message, you guide the model’s behavior, tailoring responses for more relevant, context-aware outputs.


When to use
Ideal for domain-specific tasks where customization is key.


Benefits

Save time by fine-tuning on the fly and enhance output quality without needing extensive retraining.


import openai

openai.api_key = "your-api-key"

response = openai.ChatCompletion.create(
  model="gpt-4",
  messages=[
    ,
    
  ]
)

print(response['choices'][0]['message']['content'])

At Alpha Signal, our mission is to build a sharp, engaged community focused on AI, machine learning, and cutting-edge language models, helping over 200,000 developers stay informed and ahead. We’re passionate about curating the best in AI, from top research and trending technical blogs to expert insights and tailored job opportunities. We keep you connected to the breakthroughs and discussions that matter, so you can stay in the loop without endless searching. We also work closely with partners who value the future of AI, including employers and advertisers who want to reach an audience as passionate about AI as we are.


Our partnerships are based on shared values of ethics, responsibility, and a commitment to building a better world through technology.Privacy is a priority at Alpha Signal. Our Privacy Policy clearly explains how we collect, store, and use your personal and non-personal information. By using our website, you accept these terms, which you can review on our website. This policy applies across all Alpha Signal pages, outlining your rights and how to contact us if you want to adjust the use of your information. We’re based in the United States. By using our site, you agree to be governed by U.S. laws.

Looking to promote your company, product, service, or event to 200,000+ AI developers? Let's work together.

Work with us