法国人工智能初创公司Mistral在激烈的AI竞赛中，推出其新一代旗舰模型——Mistral Large 2。这一模型的参数量达到1230亿，相较于Meta公司昨日发布的开源Llama 3.1模型的4050亿参数，Mistral Large 2在参数量上相对较少，但其在代码生成、数学推理以及多语言支持方面表现突出，官方宣称其性能接近GPT-4、Llama 3.1-405和Anthropic的Claude 3.5 Sonnet模型。
Mistral Large 2模型的一大亮点是其广泛的语言支持能力，能够处理80多种编程语言，提供强大的代码生成和数学推理能力。此外，该模型还具备128k的上下文窗口，能够处理包括中文在内的多种语言。在MMLU基准测试中，Mistral Large 2的准确度达到84.0%，展现了其在多个关键领域上的出色性能。
在开放方式上，Mistral Large 2关注非商业研究用途，提供模型权重的开放，允许第三方进行微调（fine-tune）。商业或企业用户若需使用此模型，则需向Mistral公司购买单独的许可和使用协议。
Mistral公司强调，通过专注于减少模型的幻觉问题，Large 2能够提供更为精确和有辨别力的响应，当模型对某个问题没有答案时，它会承认自己不知道，避免了编造看似合理但实际错误的答案。
Mistral Large 2的发布，标志着公司在推动AI成本效益、速度和性能发展方面取得重要进展。未来，Mistral计划继续推出更多功能，如高级函数调用和检索，以支持构建高性能的人工智能应用。
Mistral Large 2的发布，不仅为AI领域的研究和应用提供了新的工具，也展示了法国在人工智能创新领域的实力。随着更多创新模型的涌现，AI竞赛的激烈程度将进一步提升，推动整个行业向着更加智能、高效的方向发展。
英语如下：
News Title: “French Mistral Unveils 123 Billion Parameter AI Model, Challenging GPT-4 and Llama 3.1”
Keywords: AI model, parameter count, performance near
News Content:
News Title: French AI startup Mistral Launches Flagship Model Large 2, with 123 billion parameters, challenging industry giants like GPT-4 and Llama 3.1
News Body:
In a high-stakes AI race, French AI startup Mistral has unveiled its latest flagship model, Mistral Large 2, in the field of AI. This model boasts a parameter count of 123 billion, which is notably lower compared to Meta’s recently released open-source Llama 3.1 model with 4050 billion parameters. Despite the lower parameter count, Mistral Large 2 excels in code generation, mathematical reasoning, and multilingual support, with claims of performance that rivals GPT-4, Llama 3.1-405, and Anthropic’s Claude 3.5 Sonnet.
A standout feature of Mistral Large 2 is its broad language support, capable of handling over 80 programming languages and offering robust capabilities in code generation and mathematical reasoning. The model also boasts a context window of 128k, enabling it to handle multiple languages including Chinese. In the MMLU benchmark test, Mistral Large 2 achieved an accuracy of 84.0%, showcasing its exceptional performance across key domains.
Regarding accessibility, Mistral Large 2 is geared towards non-commercial research purposes, offering open access to model weights for third-party fine-tuning. For commercial or enterprise use, separate licensing and usage agreements must be purchased from Mistral.
Mistral emphasizes its focus on reducing the model’s hallucination issues, ensuring more precise and discerning responses. When the model lacks an answer to a question, it acknowledges its ignorance, avoiding the creation of seemingly plausible but actually incorrect answers.
The release of Mistral Large 2 signifies significant progress in AI in terms of cost-effectiveness, speed, and performance. Looking ahead, Mistral plans to introduce additional features such as advanced function calls and retrieval to support the development of high-performance AI applications.
The unveiling of Mistral Large 2 not only introduces a new tool for AI research and applications but also highlights France’s prowess in AI innovation. As more innovative models emerge, the intensity of the AI competition is expected to increase, driving the entire industry towards greater intelligence and efficiency.
【来源】https://www.ithome.com/0/784/032.htm
Views: 2