Meta推出了Llama 4系列AI模型，包括Scout，Maverick和Beamemoth

Meta于2025年4月5日正式推出了其最新的AI模型系列Llama 4。该版本包括四个车型：Scout，Maverick，Maverick，Embemoth和L4emoth。

Meta has unveiled its latest series of AI models, Llama 4, in a move that signals the company's persistence in the competitive landscape of large language models (LLMs).

Meta推出了其最新系列的AI模型Llama 4，此举标志着公司在大型语言模型（LLMS）竞争性景观中的持久性。

This release, which follows the previous iteration launched in 2024, includes four models: Scout, Maverick, Behemoth, and L4emoth.

此版本是在2024年推出的先前迭代之后的，其中包括四个型号：侦察兵，小牛，庞然大物和L4emoth。

Each model is designed with a focus on multimodal capabilities and computational efficiency, leveraging a mixture-of-experts (MoE) architecture.

每个模型都旨在关注多模式功能和计算效率，利用Experts（MOE）体系结构的混合物。

The MoE approach allows the models to divide tasks into subtasks handled by specialized "expert" components, enhancing performance while reducing computational costs.

MOE方法允许模型将任务分为由专业的“专家”组件处理的子任务，从而提高性能，同时降低计算成本。

Scout and Maverick are currently available through Meta's platforms and partners like Hugging Face, while Behemoth remains in training.

目前可以通过Meta的平台和Hugging Face等合作伙伴获得Scout和Maverick，而庞然大物仍在训练中。

Scout operates with 109 billion total parameters and excels in processing long-context documents with a context window of up to 10 million tokens. Maverick, boasting 400 billion parameters, is optimized for general assistant applications, including creative writing and multilingual tasks.

侦察兵以1009亿个参数运行，并在处理长期以来的上下文窗口高达1000万个令牌时出色。 Maverick拥有4000亿参数，针对一般助理应用程序进行了优化，包括创意写作和多语言任务。

Once released, Behemoth is expected to be the company's most powerful model yet, targeting more specialized STEM-related applications. L4emoth, an even larger model with 568 billion parameters, is designed for extreme efficiency, performing tasks with minimal computational resources.

一旦发布，Bememoth有望成为该公司最强大的模型，以针对更专业的STEM相关应用程序。 L4emoth是一个更大的模型，具有5680亿个参数，是为极端效率而设计的，以最少的计算资源执行任务。

Llama 4 models are trained on a massive dataset of text and code in multiple programming languages. They can perform various tasks, including

Llama 4型号在多种编程语言的大量文本和代码数据集上进行了培训。他们可以执行各种任务，包括

* Translation

* 翻译

* Writing different kinds of creative content

*编写各种创意内容

* Answering questions in an informative way

*以信息的方式回答问题

* Following instructions and completing requests thoughtfully

*按照说明和周到的请求完成

* Coding in multiple programming languages

*用多种编程语言编码

Despite impressive capabilities, the models still lag behind advanced offerings from Google and Anthropic in certain areas like reasoning.

尽管功能令人印象深刻，但这些模型仍然落后于Google的高级产品和在某些领域（例如推理）的拟人产品。

"We are continuing to work on improving our models' reasoning abilities and expect to share more updates on that front in the coming months," a Meta spokesperson stated.

一位发言人说：“我们将继续致力于提高模型的推理能力，并期望在接下来的几个月中分享有关该方面的更多更新。”

This launch marks the beginning of a new phase for the Llama ecosystem.

此发布标志着Llama生态系统的新阶段的开始。

* The models will be supported by a vibrant community of researchers and developers on the Llama GitHub repository.

*这些模型将得到充满活力的研究人员和开发人员社区的支持。

* The company also plans to release more smaller derivative models from Llama 4 to encourage experimentation and innovation.

*该公司还计划从Llama 4释放更多较小的衍生模型，以鼓励实验和创新。

* Furthermore, Meta hopes to expand the availability of Llama 4's multimodal features to more users and regions in the future.

*此外，Meta希望将来将Llama 4的多模式功能扩展到更多的用户和地区。

This move comes as Big Tech companies face increasing pressure to balance technological advancement with ethical and legal considerations, especially regarding data privacy and the potential misuse of AI.

这一举动是因为大型科技公司面临越来越多的压力，以平衡技术进步与道德和法律方面的考虑，尤其是在数据隐私以及AI的潜在滥用方面。

In Europe, for instance, Meta's ability to deploy models like Llama 4 is subject to licensing restrictions due to EU regulations. Additionally, companies with over 700 million monthly active users require special licenses to operate in the region.

例如，在欧洲，Meta的部署模型（如Llama 4）的能力受到欧盟法规的许可限制。此外，拥有超过7亿个活跃用户的公司要求特殊许可证在该地区运营。

Llama 4 represents Meta's response to escalating competition in the LLM domain. While internal benchmarks suggest improvements over some competitors in specific tasks, the models still lag behind advanced offerings from Google and Anthropic in certain areas like reasoning.

Llama 4代表了Meta对LLM领域中竞争不断升级的反应。虽然内部基准测试表明，在特定任务中，对某些竞争对手进行了改进，但这些模型仍然落后于Google的高级产品和在某些领域（例如推理）的拟人产品。

"We are continuing to work on improving our models' reasoning abilities and expect to share more updates on that front in the coming months," a Meta spokesperson noted.

一位发言人指出：“我们将继续致力于提高模型的推理能力，并期望在接下来的几个月中分享有关该方面的更多更新。”

This launch marks the beginning of a new phase for the Llama ecosystem. The models will be supported by a vibrant community of researchers and developers on the Llama GitHub repository.

此发布标志着Llama生态系统的新阶段的开始。这些模型将得到充满活力的研究人员和开发人员社区的支持。

The company also plans to release more smaller derivative models from Llama 4 to encourage experimentation and innovation.

该公司还计划从Llama 4释放更多较小的衍生模型，以鼓励实验和创新。

Furthermore, Meta hopes to expand the availability of Llama 4's multimodal features to more users and regions in the future.

此外，Meta希望将来将Llama 4的多模式功能扩展到更多的用户和地区。

This move comes as Big Tech companies face increasing pressure to balance technological advancement with ethical and legal considerations, especially regarding data privacy and the potential misuse of AI.

这一举动是因为大型科技公司面临越来越多的压力，以平衡技术进步与道德和法律方面的考虑，尤其是在数据隐私以及AI的潜在滥用方面。

In Europe, for instance, Meta's ability to deploy models like Llama 4 is subject to licensing restrictions due to EU regulations. Additionally, companies with over 700 million monthly active users require special licenses to operate in the region.

例如，在欧洲，Meta的部署模型（如Llama 4）的能力受到欧盟法规的许可限制。此外，拥有超过7亿个活跃用户的公司要求特殊许可证在该地区运营。