Skip to content
TheChatArchiveThe Chat Archive
Perplexitysonar· PerplexityHugging Face LMArena Search Arena 24K

Llama4

From a Chatbot Arena battleThe same person sent this prompt to two anonymous models at once. Compare it with gpt-4o-search-preview-high’s answer.
  1. User

    Llama4

  2. Perplexity Assistant

    Meta has launched Llama 4, its latest suite of AI models, marking a significant advancement in open-weight, multimodal artificial intelligence. The release includes three models—Llama 4 Scout, Llama 4 Maverick, and the upcoming Llama 4 Behemoth—each leveraging a mixture-of-experts (MoE) architecture for improved efficiency and performance. These models are designed to handle text, image, and video inputs, with capabilities spanning multilingual tasks, long-context processing, and advanced reasoning.

    Key Models in the Llama 4 Series

    Llama 4 Scout

    • Parameters: 17 billion active, 109 billion total (16 experts)
    • Context Window: 10 million tokens (industry-leading)
    • Use Cases: Document summarization, codebase analysis, and multimodal tasks
    • Hardware: Runs on a single Nvidia H100 GPU with Int4 quantization [[1]][[2]][[5]]

    Llama 4 Maverick

    • Parameters: 17 billion active, 400 billion total (128 experts)
    • Context Window: 1 million tokens
    • Use Cases: Creative writing, multilingual chat, and image understanding
    • Performance: Competes with GPT-4o and Gemini 2.0 in benchmarks [[1]][[3]][[6]]

    Llama 4 Behemoth (Upcoming)

    • Parameters: 288 billion active, ~2 trillion total (16 experts)
    • Training Status: Still in progress, but early benchmarks show it outperforms GPT-4.5 and Claude 3.7 in STEM tasks [[1]][[4]][[8]]

    Innovations and Architecture

    • Mixture of Experts (MoE): Distributes computational tasks across specialized sub-models, improving efficiency. For example, Maverick uses only 17B of its 400B parameters per query [[1]][[3]][[6]].
    • Native Multimodality: Early fusion integrates text and vision tokens into a unified backbone, enabling joint training on diverse data [[2]][[6]].
    • Training Efficiency: FP8 precision and MetaP hyper-parameter tuning reduce costs while maintaining quality [[2]][[6]].

    Availability and Licensing

    • Access: Scout and Maverick are available via Meta’s partners (Hugging Face, Groq) and integrated into Meta AI across WhatsApp, Messenger, and Instagram in 40 countries [[1]][[3]][[7]].
    • Restrictions: EU users are barred due to regulatory concerns, and large enterprises (>700M MAU) require special licensing [[1]][[5]].

    Performance Highlights

    • Scout: Excels in long-context tasks (e.g., 10M tokens for parsing entire codebases) and matches Gemini 2.0 in benchmarks [[2]][[4]].
    • Maverick: Outperforms GPT-4o in coding and reasoning, with lower operational costs than Llama 3.3 70B [[3]][[6]].
    • Behemoth: Preliminary results show superiority in MATH-500 and GPQA Diamond benchmarks [[2]][[8]].

    Developer Integration

    • Groq Pricing: Scout costs 0.11/Minputtokens,Maverick0.11/M input tokens, Maverick 0.50/M input tokens [[3]].
    • Hugging Face: Supports transformers and TGI for easy fine-tuning and deployment [[7]].

    Meta positions Llama 4 as a foundational shift toward open, scalable AI, with Behemoth expected to further push boundaries upon release. The models balance performance and accessibility, though licensing and regional limitations may impact adoption.

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.