0
0
Alibaba Unveils Qwen2: The Next Generation of AI Models
Alibaba, the Chinese e-commerce giant, continues to make strides in the AI field with the release of its latest AI model, Qwen2. This new model, developed by Alibaba Cloud, is part of the Tongyi Qianwen (Qwen) series, which includes the Tongyi Qianwen LLM, Qwen-VL for vision AI, and Qwen-Audio.
The Qwen series is known for being pre-trained on multilingual data across various industries, with the Qwen-72B model being the most powerful in the lineup. Trained on a massive 3 trillion tokens of data, Qwen2 surpasses Meta’s Llama-2 variant, which is based on 2 trillion tokens. While Llama-3 is in the process of digesting 15 trillion tokens, Qwen2 stands out for handling 128K tokens of context, comparable to GPT-4o from OpenAI.
Qwen2: Setting a New Standard in AI
According to the Qwen team’s recent blog post, Qwen2 outperforms Meta’s LLama3 in key synthetic benchmarks, positioning itself as the best open-source model available. However, the Elo Arena ranks Qwen2-72B-Instruct slightly below other models but as the second most favored open-source LLM among human testers.
Offered in five different sizes, ranging from 0.5 billion to 72 billion parameters, Qwen2 brings significant improvements across various domains. Trained on data in 27 languages, including English, Chinese, German, French, Spanish, Italian, and Russian, Qwen2 showcases its multilingual capabilities.
The Qwen team stated on HuggingFace that Qwen2-72B and its instruction-tuned models maintain the original Qianwen license, while other models now use the standard Apache 2.0 license, aiming to accelerate open-source AI.
Advancing Language Understanding with Qwen2
Qwen2 exhibits a remarkable understanding of long contexts, excelling in information extraction tasks and passing the “Needle in a Haystack” test with flying colors. The model’s performance remains consistent even with prolonged interactions, a notable feat in the AI realm.
Moreover, Qwen2 displays proficiency in system prompts, adapting its responses based on specified conditions. The model’s upcoming upgrade will introduce multimodality, enabling it to understand both vision and audio information.
For those interested in testing Qwen2, the model is available for online testing via HuggingFace Spaces. Additionally, users with sufficient computing power can download the weights for free from HuggingFace’s platform.
Embracing Open-Source AI with Qwen2
Qwen2 emerges as a compelling choice for enthusiasts of open-source AI, offering a larger token context window than many models in the market. With its flexible license, collaborative fine-tuning efforts could further enhance Qwen2’s capabilities, fostering a more inclusive and unbiased AI environment.
Image/Photo credit: source url
