软餐获悉，Anthropic 发布了 Claude 3 语言模型，据称其表现优于 GPT-4 和 Gemini。Anthropic 推出的第三代 AI 模型包括三个版本：Haiku、Sonnet 和 Opus。Claude 3 可以对长达 150,000 个词的文本进行摘要，并且在竞争的语言模型中表现更好。AI 初创公司 Anthropic 表示，在测试中，Claude 3 的 Opus 模型在包括大学水平推理、编程和数学等方面表现优于竞争对手 OpenAI 和谷歌的 GPT-4 和 Gemini 1.0 Ultra。Opus 是最先进且最昂贵的模型，Haiku 是最小的版本，被称为 “最快且成本效益最高”。该模型应能够在不到三秒钟内阅读包含图表和图形的简明研究报告。
这是 Claude 首次提供多模态支持。Claude 3 可以同时处理文本和图像输入，并且应能够对长达 150,000 个词的文本进行摘要，是其前身的两倍。Anthropic 表示，与 Claude 2.1 相比，第三代语言模型能够更好地理解上下文，处理更长的指令，并且更准确。据称，Sonnet 是中等模型，速度是上一代的两倍，在 “需要快速响应的任务” 方面表现出色。
Anthropic 的 Claude 3 模型根据一份文件使用了来自 2023 年 8 月以后的公开信息和第三方的非公开数据进行训练。语言模型的训练使用了亚马逊网络服务和谷歌云的硬件。亚马逊和谷歌都对这家初创公司进行了投资。
Opus 和 Sonnet 从今天开始可用，Haiku 的发布将在不久之后。Anthropic 对 Claude 3 Opus 处理的每一百万个标记收取 15 美元的费用。相比之下，OpenAI 对 GPT-4 Turbo 处理的每一百万个标记收费 10 美元。Sonnet 和 Haiku 分别为每一百万个标记收取 3 美元和 0.25 美元。Claude 还有一个限量版的免费版本。
Anthropic has released the Claude 3 language model, which is claimed to outperform GPT-4 and Gemini. The third-generation AI model introduced by Anthropic includes three versions: Haiku, Sonnet, and Opus. Claude 3 is capable of summarizing texts of up to 150,000 words and performs better than competing language models.
AI startup Anthropic states that in tests, Claude 3’s Opus model performed better than GPT-4 and Gemini 1.0 Ultra from competitors OpenAI and Google in areas such as university-level reasoning, programming, and mathematics. Opus is the most advanced and expensive model, while Haiku is the smallest version, known for being “the fastest and most cost-effective.” The model should be able to read a concise research report with diagrams and graphics “in less than three seconds.”
This is the first time Claude provides multimodal support. Claude 3 can handle both text and image inputs and is expected to summarize texts of up to 150,000 words, twice as much as its predecessor. According to Anthropic, the third generation of its language model can better understand context, process longer instructions, and be more accurate than Claude 2.1. Sonnet, the medium-sized model, is reportedly twice as fast as the previous generation and excels in “tasks that require rapid responses.”
Anthropic’s Claude 3 models were trained on publicly available information from August 2023 onwards, as well as non-public third-party data, as stated in a paper. The training of the language models utilized the hardware of Amazon Web Services and Google Cloud. Both Amazon and Google have invested in the startup.
Opus and Sonnet are available starting today, while the release of Haiku will follow “soon.” Anthropic charges $15 for every million tokens processed by Claude 3 Opus. In comparison, OpenAI charges $10 for every million tokens processed by GPT-4 Turbo. For Sonnet and Haiku, the respective costs are $3 and $0.25 for 1 million tokens. There is also a limited free version of Claude available, although Anthropic’s language models are currently not officially available in the Netherlands and Belgium.
微信扫一扫