Alibaba Cloud has released the open-source LLM model from the Qwen 2 lineage, introducing two additional models: Qwen2-Math and Qwen2-Audio.
Qwen2-Math is a model designed to further enhance the capabilities of Qwen2, utilizing a high-quality mathematical data set including textbooks, various code snippets, exam sets, and synthesized data from Qwen2 itself. The standout feature of this model is its superior performance in mathematics tests such as GSM8K, MATH, or MMLU-STEM, outperforming closed models like GPT-4o or Gemini.
Qwen2-Math is currently only compatible with the English language and comes in three sizes: 1.5B, 7B, and 72B, available under the Apache 2.0 license. The team has announced plans to release a Chinese language version in the near future.
On the other hand, Qwen2-Audio is a model specifically tailored for direct voice chat, allowing for the input of voice alone or voice with accompanying text. Text inputs can include commands related to sound analysis, supporting 8 languages including Chinese, English, Cantonese, French, Italian, Spanish, German, and Japanese.
The architecture of Qwen2-Audio features a new encoder to support direct sound, along with model training on various sound-related data sets. The Qwen2-Audio model is available in a single size of 7B, with a separate instruct version.
TLDR: Alibaba Cloud introduces the Qwen 2 lineage with the open-source LLM model, offering specialized models like Qwen2-Math for mathematics and Qwen2-Audio for direct voice chat applications in multiple languages.
Leave a Comment