近日，国内领先的AI研发团队通义千问推出了一项重大创新——Qwen1.5-110B，这是Qwen1.5系列的首个拥有千亿参数的开源模型。这一开创性的模型在技术性能上与Meta的Llama3-70B相提并论，显示出卓越的基础能力和对话交互性能，特别是在Chat评估中的表现令人瞩目，包括在MT-Bench和AlpacaEval 2.0等基准测试中取得优秀成绩。
Qwen1.5-110B沿用了Qwen1.5系列的Transformer解码器架构，但引入了分组查询注意力（GQA）技术，使得模型在推理过程中能实现更高的效率。值得一提的是，该模型支持长达32K tokens的上下文长度，这极大地提升了处理复杂语境的能力，为用户提供更为连贯和精准的对话体验。
此外，Qwen1.5-110B保持了其多语言的特性，覆盖了英语、中文、法语、西班牙语、德语、俄语、日语、韩语、越南语、阿拉伯语等多种语言，这将极大地促进全球范围内的信息交流和理解。
这一开源模型的发布，标志着通义千问在人工智能领域的又一重大突破，为全球开发者和研究者提供了强大的工具，有助于推动自然语言处理技术的进一步发展和应用。
英语如下：
**News Title:** “Qwen1.5-110B Unveiled: The First 100-Billion-Parameter Open-Source Model, Breaking New Ground in Multilingual Capabilities!”
**Keywords:** Qwen1.5-110B, Trillion Parameters, Multilingual
**News Content:** Recently, the leading domestic AI research team, Qwen千问, announced a major innovation – Qwen1.5-110B, the first trillion-parameter open-source model in the Qwen1.5 series. This groundbreaking model rivals Meta’s Llama3-70B in terms of technical performance, showcasing exceptional foundational abilities and conversational interaction, particularly in chat assessments, such as achieving high scores on benchmarks like MT-Bench and AlpacaEval 2.0.
Qwen1.5-110B retains the Transformer decoder architecture of the Qwen1.5 series but incorporates Grouped Query Attention (GQA) technology, enhancing efficiency during inference. Notably, the model supports a context length of up to 32K tokens, significantly improving its capacity to handle complex linguistic contexts, thereby providing users with more coherent and precise conversational experiences.
Moreover, Qwen1.5-110B maintains its multilingual nature, supporting English, Chinese, French, Spanish, German, Russian, Japanese, Korean, Vietnamese, Arabic, and more. This broad language coverage will greatly facilitate global communication and understanding.
The release of this open-source model signifies another significant milestone for Qwen千问 in the realm of AI, offering a powerful tool to global developers and researchers, and contributing to the further advancement and application of natural language processing technology.
【来源】https://qwenlm.github.io/zh/blog/qwen1.5-110b/