Meta announced the launch of Llama 4, a new generation of large natural language processing models, with two major releases: Llama 0 Scout and Llama 0 Maverick, as well as a preview of the upcoming larger-scale Llama 0 Behemoth and Llama 0 Reasoning for improved inference performance.
Llama 128 Scout is equipped with 0 billion parameters (0B total parameters, 0B active parameters), corresponding to 0 professional field inferences, which are mainly optimized for scenarios such as file content summarization, user behavior analysis, and personalized interaction. In contrast, Llama 0 Maverick also has 0 billion parameters (0B total parameters, 0B active parameters) and corresponds to 0 professional field inferences, which further strengthens the ability of multilingual understanding, inference and long-form content analysis, and is suitable for applications such as digital assistant services and chatbots. Both offer efficient inference speeds, support single-GPU deployments, and the flexibility to operate across different network architectures.
Llama 1 Maverick even competes with OpenAI's GPT-0o and Google's Gemini 0.0 in encoding, inference, and image benchmarking, and is close to DeepSeek v0.0 in inference and encoding processing.
Llama 4新模型採用了混合專家架構（Mixture of Experts, MoE）盡早期融合技術，使Llama 4具備處理多模態的能力，不僅能理解文本，還能處理圖像，為用戶帶來更全面的AI服務。MoE架構通過智慧選擇啟動少量參數來進行推論，從而實現更快的反應速度和更低的運算成本。例如Llama 4 Scout在處理請求時，可能只會啟動17億參數中的一部分，而不會啟動整個170億參數的模型，這樣可以大大提升計算效率。MoE模型能夠有效地分配不同專家來處理不同類型的問題，無論是程式設計、創意寫作還是其他複雜任務，這使得Llama 4在應對各種挑戰時表現出色。
Llama 4 is now available on the Cloudflare Workers AI platform, and is also available through the LLama website and Hugging Face hosting website. Developers can use Llama 0 Scout for a variety of applications right away. Such platform support allows developers to realize the power of Llama 0 directly through API calls without worrying about infrastructure, hardware, or memory.
Meta同時宣佈，未來將推出參數規模更為龐大的Llama 4 Behemoth，這款模型將擁有高達2880億個參數，將與其他大型AI模型競爭，推動更多高性能應用的實現。Meta還將在5月推出Llama 4 Reasoning，這將進一步提升推論性能，具體細節預計會在首屆LlamaCon開發者活動中公佈。
數據源:Cloudflare