Image source: generated by Unbounded AI

The large model circle these two days has been very lively. On the other hand, DeepSeek has attracted countless attention with its low-cost, high-performance and outstanding performance, and various technology experts have praised it.

On the other hand, various strange "emojis" have become popular on social media and technology forums, such as "QwQ" and "QVQ". People who don't know what they mean may look confused, but they know what they are doing. Everyone knows that this refers to the series of models of Alibaba Tongyi Qianwen Open Source.

In September this year, Alibaba released Tongyi Qianwen’s new generation open source model Qwen2.5 series, launching versions of different sizes from 0.5B to 72B parameters in one go, covering various application needs, successfully becoming one of the It ranks among the world's top open source models, with multi-modal and multi-language capabilities, and has become a favorite of many companies and developers. Not only that, the Qwen team has also been active recently, and has successively open sourced several bright AI models, which continues to attract attention.

The naming style of the Qwen team is also quite "abstract": QVQ, "staring with two eyes"; QwQ, like an accidental scratch on the keyboard. What code world emoji is this? It seems that the technical experts secretly hide a little "skin" apart from their serious scientific research. Perhaps, Qwen's naming philosophy is: the name is arbitrary, but the last word is that it can be typed?

When it comes to generative AI, everyone’s eyes always seem to be inseparable from the technology giants on the other side of the ocean: OpenAI, Google, and Meta. But recently, some figures from the East have appeared frequently in the AI ​​circle, such as DeepSeek and Alibaba’s Tongyi Qianwen Qwen. They are conquering the major AI model performance lists and becoming the focus.

You know, when we talked about domestic large models in the past, we always felt that there was a bit of a "catcher" shadow. But now China's open source power is using its strength to prove that it is no longer a spectator, but a role that can compete with giants such as OpenAI and Meta.

Hugging Face’s 2024 annual inventory data is very interesting: Qwen2.5-1.5B-Instruct’s downloads accounted for 26.6%, crushing star open source models such as Llama 3 and Gemma. Although the number of downloads cannot be completely equated to "the strongest strength", it is definitely a hard indicator of popularity.

The ultra-high download volume of Qwen2.5-1.5B-Instruct proves its wide application and high popularity at the current time point, and also reflects that the open source large models developed by Chinese companies are on the international stage. Showing increasing influence, in fact, the most downloaded open source model on the Hugging Face platform in 2023 is also from the Chinese community and is BGE-base of Zhiyuan Research Institute.

For QwenForeign netizens are also happy to see the outstanding performance, and they have even started to joke: Zuckerberg may be secretly watching you use Qwen instead of Llama. ”

1. Christmas gift package QvQ, the first open source multi-modal reasoning model

Netizens started to use the Christmas gift given by the Qwen team: QVQ-72B-Preview, which is a new model that can analyze images and perform inference. Open source model. Although still in the experimental stage, preliminary tests show that it performs quite well in visual reasoning tasks.

QVQ solves problems by thinking step by step, similar to OpenAI's o1 or Google's Flash. Models such as Thinking that "think step by step". The user throws it a picture and some instructions, and the system analyzes the information, takes time to reflect when necessary, and provides an answer with a confidence score for each prediction.

In terms of underlying architecture, QVQ-72B-Preview is built on Qwen’s existing visual language model Qwen2-VL-72B, and adds the ability to think and reason, making it the first open source model of its kind.

p>

Developer on Macbook Running QVQ on Pro

In the benchmark test, the open source QVQ comprehensively surpassed its predecessor Qwen2-VL-72B-Instruct and reached an accuracy level similar to closed source models such as OpenAI's o1 and Claude 3.5 Sonnet. .

QVQ-72B-Preview Benchmark test results

In the actual test, a netizen took a photo of the New York subway to test QVQ and asked "If I want to go to Chinatown, should I get off at this station?" "The user finally said that the model made correct judgments and conducted effective reasoning based on the problem.

2. Open source reasoning model praised by Tao Zhexuan QwQ

Moving forward, on November 28, 2024, the Qwen team also open sourced an AI model focusing on reasoning capabilities. QwQ-32B-Preview.

This is the first open source inference model released by the Qwen team, aiming to enhance AI inference capabilities. Although the parameter size is only 32B, it has excellent performance in GPQA, AIME, MATH-500 and LiveCodeBench. Waiting for multiple reviews, QwQ All achieved good results, even surpassing o1 in some tests. QwQ has the ability to deeply introspect, question its own assumptions and engage in thoughtful self-dialogue to solve complex problems.

Although QwQ is still in the experimental stage, its powerful analytical capabilities and unique reasoning methods have attracted a lot of attention. Even the great figure in mathematics, Terence Tao, publicly praised it, saying that its performance surpassed previous ones. All open source models.

In the AIMO (AI Mathematics Olympiad) challenge, the Qwen series models have also become one of the most commonly used models by contestants, ranking among the top three.

The top three most used models in AIMO are all Qwen

“With open weights, low prices, and outstanding basic capabilities, who wouldn’t like such inference models?”

3. Qwen2.5-Coder: The "code leader" in the open source world?

The release of the Qwen2.5 series, especially the appearance of Qwen2.5-Coder, triggered a wave of discussions in the AI ​​circle. Despite its relatively small model size, Qwen 2.5 Coder32B is still comparable to leading-edge models in programming benchmarks such as HumanEval.

Some overseas technology bloggers complain that now everyone seems to be focusing on the developments of giants such as OpenAI, Google, and Anthropic, but ignore the "ruthless character" of Qwen. It is the first open weight model that can compete head-on with Claude Sonnet and GPT-4o, and can also be run locally on your computer. This is not just about good-looking results, many people who have experienced it say it "smells really good." In contrast, although DeepSeek's model is also very good, it is too large and difficult to run locally (deepseek v3 has not been released yet). The emergence of Qwen2.5-Coder is definitely big news for the open source community. What's more conscientious is that Alibaba also made the technical report completely public, without hiding it, and shared the results with the community.

There are also developers who have built an AI video editor Video Composer based on Qwen2.5-Coder. Users can drag and drop materials (such as pictures, videos and audios) and use natural language to make Qwen2. 5-Coder generates new videos (based on FFMPEG technology).

4. Meet diverse needs and global Qwen

Another big advantage of Qwen is its "approachability". The Qwen2.5 series is not just for technical experts or large enterprises, it is designed to be easily used by a wide range of users. From the 50 million parameter version suitable for resource-constrained devices to the 72 billion parameter version required for enterprise-level applications, a wealth of choices are provided to meet different needs..

In Japan, Alibaba Cloud has partnered with University of Tokyo startup Lightblue to improve its Japanese large language models (LLMs). Lightblue uses Alibaba Cloud's architecture and Qwen LLM technology to optimize the model, thereby improving the accuracy of East Asian languages.

Well-known investor and former CTOBalaji Srinivasan of Coinbase also publicly recognized Qwen’s multi-modal and multi-language capabilities.

Qwen’s various models can now be easily accessed by engineers around the world. What’s even more rare is that Qwen performs well in processing multiple languages. Even some “small languages” with relatively small global AI training data, such as Burmese, Bengali and Urdu, can handle it. In contrast, Meta’s open source AI model Llama is mainly targeted at English applications.

Many Japanese developers are carefully studying Qwen2.5’s technical report

5. The rise of China’s AI open source power

The rise of Chinese AI models such as Qwen has provided domestic companies with more choices and possibilities. In the current international environment, its significance is even more prominent. More importantly, they are not just a "spare tire" option, but are proving their ability to compete with top American technology.

The significance of Qwen is not only reflected in technology, but also the openness and collaboration represented behind it, which shows that China has not fallen behind in the field of AI, but has shown strong competitiveness through open source. It turns out that the so-called GPU limitations have not hindered the development of AI in China. If this momentum is maintained, China is likely to occupy a more important position in the LLM market. When the open source model is more open than Meta (which releases the model with a special Llama research license), and when everyone can use the open source model with the same or even better performance, who wouldn't be happy to use it?

CNBC also recently published an article stating that China has made significant progress in LLM, and models such as Qwen and DeepSeek have surpassed American competitors in some aspects. Chinese companies are actively embracing the open source model and promoting the development and application of AI technology to promote innovation and expand global influence. The article believes that China is rapidly rising in the field of AI, its AI models have become quite internationally competitive, and it is working hard to build an independent and controllable AI ecosystem.

Hugging Face CEO Clem even mentioned in his 2025 AI predictions that China will begin to lead the artificial intelligence competition, mainly due to its leading position in the open source artificial intelligence competition.

Sam Altman recently "lamented": It is relatively easy to copy, but to do something new and riskyLove is extremely difficult. But he also said successful individual researchers deserved the credit because it was "the coolest thing in the world." In the comment area, Vaibhav Srivastav responded that public sharing should not be ignored, and named Qwen and the DeepSeek team, who also deserve applause.

An open mind, coupled with an emphasis on engineering practice, is accelerating the development of China’s AI industry. China's AI industry, which was once thought to be hampered by semiconductor limitations and limited computing power, is proving to the world, represented by the open source model, that it has the ability to compete with the world's top levels and create greater value on a global scale. .