back to top
Saturday, June 8, 2024
HomeCryptoAlibaba launches Qwen2 AI to challenge Meta and OpenAI

Alibaba launches Qwen2 AI to challenge Meta and OpenAI

Alibaba, the Chinese e-commerce giant, is a major player in China’s AI sector. Today, they announced the launch of their latest AI model, Qwen2 – and by some comparative measures, it’s the best open source option available today.

Developed by Alibaba Cloud, Qwen2 is the next generation of the company’s Tongyi Qianwen (Qwen) model line, which includes the large language model Tongyi Qianwen LLM (also known as Qwen), the image AI model Qwen -VL and Qwen-Audio.

Qwen is pre-trained on multilingual data covering a variety of industries and domains, with the Qwen-72B being the most powerful model in the family. It was trained on 3 trillion tokens of data. In comparison, Meta’s strongest variant of Llama-2 is based on 2 trillion tokens. However, Llama-3 is in the process of processing 15 trillion tokens.

According to a post blog recently released by the Qwen team, Qwen2 can handle 128k context tokens – equivalent to GPT-4o from OpenAI. Qwen2 also outperformed Meta’s LLama3 in most important composite metrics, the team claims, making it the best open source model available.

However, it’s worth noting that Elo Arena independently rates the Qwen2-72B-Instruct slightly higher than the GPT-4-0314 but below the LLama3 70B and GPT-4-0125-preview, making it the LLM model Open source is the second most popular among testers so far.

Qwen2 performs better than Llama3, Mixtral and Qwen1.5 in synthetic benchmarks | Image: Alibaba Cloud

Qwen2 is available in five sizes ranging from 0.5 billion to 72 billion parameters, and this release brings significant improvements in various areas of expertise. Additionally, the models were trained with data in 27 more languages ​​than the previous version, including German, French, Spanish, Italian and Russian, in addition to English and Chinese .

“Compared to the most advanced open source language models, including the previously released Qwen1.5, Qwen2 generally surpasses most open source models and demonstrates competitiveness with proprietary models on a range of metrics targeting language comprehension, language production, multilingualism, programming, mathematics and reasoning,” the Qwen team states on the model’s official page on HuggingFace.

Qwen2 models also show impressive ability to understand long contexts. The Qwen2-72B-Instruct can handle information extraction tasks anywhere in its massive context without error, and it passed the “Needle in a Haystack” test almost flawlessly. . This is important, because traditionally, a model’s performance starts to degrade the more we interact with it.

Qwen2 performs amazingly on the “Needle in a Haystack” test | Photo: Alibaba Cloud

With this release, the Qwen team has also changed the license for its models. While the Qwen2-72B and its instruction adaptation models continue to use the original Qianwen license, all other models have adopted the Apache 2.0 license, a standard in the open source software world.

“In the near future, we will continue to open code new models to accelerate open source AI,” Alibaba Cloud said in an official blog post.

Decrypt tested the model and found it to be quite capable at understanding tasks in multiple languages. This model is also subject to censorship, especially in topics considered sensitive in China. This appears to be consistent with Alibaba’s claim that Qwen2 is the model least likely to deliver unsafe results – whether illegal activity, fraud, pornography, and privacy violations. private – no matter what language.

ChatGPT’s answer to the sensitive question: “Is Taiwan a country?”

Additionally, it has a good understanding of system prompts, which means that the conditions applied will have a stronger influence on its answers. For example, when asked to play the role of a helpful assistant with knowledge of the law versus a knowledgeable lawyer who always answers based on the law, the responses showed a big difference. It provides similar advice to GPT-4o, but is more concise.

The next model upgrade will bring multimodality to the Qwen2 LLM, which can consolidate all families into one robust model, the team said. “In addition, we extend Qwen2 language models to multimodality, capable of understanding both visual and audio information,” they added.

ChatGPT’s answer to: “A neighbor insulted me”

Qwen is available for testing online via HuggingFace Spaces. Those with enough computing power to run it locally can download weights for free via HuggingFace.

The Qwen2 model can be a great alternative for those who want to bet on open source AI. It has a larger context window than most other models, making it even more powerful than Meta’s LLama3. Additionally, thanks to its license, tweaked versions shared by others can improve it, increase scores and overcome bias.

*Artificial General Intelligence (AGI) is a form of AI capable of performing all intellectual tasks that humans can do. Unlike narrow AI (ANI), AGI has the ability to understand, learn and apply knowledge in many different fields. AGI can learn on its own from experience and new data without the need for constant human intervention. It can adapt to new situations and problems that have never been encountered before. AGI is considered the ultimate goal of AI research, but currently it is still at least 10 years away from development. AGI is causing many security concerns and potential risks to humanity.

Thach Sanh

According to Decrypt

Mark Tyson
Mark Tyson
Freelance News Writer. Always interested in the way in which technology can change people's lives, and that is why I also advise individuals and companies when it comes to adopting all the advances in Apple devices and services.


Please enter your comment!
Please enter your name here