AI Agent Hub
Back to models
Llama-4-Scout-17B-16E-Instruct logo

Llama-4-Scout-17B-16E-Instruct

Multimodal meta-llama Released 2025-04-05
10.0 / 100 109.0B params 10M context Open Source

About this model

Llama-4-Scout-17B-16E-Instruct is Meta's instruction-tuned, natively multimodal mixture-of-experts model released in April 2025. It activates 17 billion parameters per token from a total of 109 billion parameters across 16 routed experts plus a shared expert, and uses early fusion to jointly process text and images without a separate vision encoder. The model was pretrained on roughly 40 trillion tokens with a knowledge cutoff of August 2024, and supports twelve languages including English, French, German, Spanish, Portuguese, Italian, Hindi, and Arabic.

Benchmark Scores

GPQA
46.46
MBPP
67.8
MMLU-Pro
74.3

Technical Specs

  • Parameters: 109.0B
  • Architecture: MoE Transformer
  • Context Window: 10,000,000 tokens
  • Input Modalities: text, image

Hardware Requirements

  • VRAM: 80.0 GB
  • Compute: Single NVIDIA H100 80GB with on-the-fly INT4 quantization; ~218GB VRAM for full BF16 weights