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

Meta於2025年4月5日正式推出了其最新的AI模型系列Llama 4。該版本包括四個車型：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，此舉標誌著公司在大型語言模型（LLMS）競爭性景觀中的持久性。

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後的，其中包括四個型號：偵察兵，小牛，龐然大物和L4emoth。

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

每個模型都旨在關注多模式功能和計算效率，利用Experts（MOE）體系結構的混合物。

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

MOE方法允許模型將任務分為由專業的“專家”組件處理的子任務，從而提高性能，同時降低計算成本。

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

目前可以通過Meta的平台和Hugging Face等合作夥伴獲得Scout和Maverick，而龐然大物仍在訓練中。

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.

偵察兵以1009億個參數運行，並在處理長期以來的上下文窗口高達1000萬個令牌時出色。 Maverick擁有4000億參數，針對一般助理應用程序進行了優化，包括創意寫作和多語言任務。

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.

一旦發布，Bememoth有望成為該公司最強大的模型，以針對更專業的STEM相關應用程序。 L4emoth是一個更大的模型，具有5680億個參數，是為極端效率而設計的，以最少的計算資源執行任務。

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型號在多種編程語言的大量文本和代碼數據集上進行了培訓。他們可以執行各種任務，包括

* Translation

* 翻譯

* Writing different kinds of creative content

*編寫各種創意內容

* Answering questions in an informative way

*以信息的方式回答問題

* Following instructions and completing requests thoughtfully

*按照說明和周到的請求完成

* Coding in multiple programming languages

*用多種編程語言編碼

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

儘管功能令人印象深刻，但這些模型仍然落後於Google的高級產品和在某些領域（例如推理）的擬人產品。

"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.

一位發言人說：“我們將繼續致力於提高模型的推理能力，並期望在接下來的幾個月中分享有關該方面的更多更新。”

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

此發布標誌著Llama生態系統的新階段的開始。

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

*這些模型將得到充滿活力的研究人員和開發人員社區的支持。

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

*該公司還計劃從Llama 4釋放更多較小的衍生模型，以鼓勵實驗和創新。

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

*此外，Meta希望將來將Llama 4的多模式功能擴展到更多的用戶和地區。

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.

這一舉動是因為大型科技公司面臨越來越多的壓力，以平衡技術進步與道德和法律方面的考慮，尤其是在數據隱私以及AI的潛在濫用方面。

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，在歐洲，Meta的部署模型（如Llama 4）的能力受到歐盟法規的許可限制。此外，擁有超過7億個活躍用戶的公司要求特殊許可證在該地區運營。

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對LLM領域中競爭不斷升級的反應。雖然內部基準測試表明，在特定任務中，對某些競爭對手進行了改進，但這些模型仍然落後於Google的高級產品和在某些領域（例如推理）的擬人產品。

"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.

一位發言人指出：“我們將繼續致力於提高模型的推理能力，並期望在接下來的幾個月中分享有關該方面的更多更新。”

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.

此發布標誌著Llama生態系統的新階段的開始。這些模型將得到充滿活力的研究人員和開發人員社區的支持。

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

該公司還計劃從Llama 4釋放更多較小的衍生模型，以鼓勵實驗和創新。

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

此外，Meta希望將來將Llama 4的多模式功能擴展到更多的用戶和地區。

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.

這一舉動是因為大型科技公司面臨越來越多的壓力，以平衡技術進步與道德和法律方面的考慮，尤其是在數據隱私以及AI的潛在濫用方面。

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，在歐洲，Meta的部署模型（如Llama 4）的能力受到歐盟法規的許可限制。此外，擁有超過7億個活躍用戶的公司要求特殊許可證在該地區運營。