LlamaMoELlama 4 Community License

Llama 4 Scout (MoE)

Llama 4 Scout is Meta's efficiency-focused MoE model. With 109B total parameters but only 17B active per token, it delivers performance comparable to dense 70B models at roughly half the inference FLOPs. However, all 109B parameters must be

109.0B

Parameters

17.0B

Active

256K

Max Context

MoE

Architecture

Apr 5, 2025

Released

Text + Vision

Modality

About Llama 4 Scout (MoE)

Llama 4 Scout is Meta's efficiency-focused MoE model. With 109B total parameters but only 17B active per token, it delivers performance comparable to dense 70B models at roughly half the inference FLOPs. However, all 109B parameters must be loaded into VRAM (~55 GB at Q4_K_M), making it a high-memory-requirement model despite its efficient compute profile. Supports 256K context and multimodal vision input. Best suited for servers or high-end workstations with 48 GB+ GPUs.

ReasoningCodeMultimodalAgentic

Technical Specifications

Total Parameters109.0B
Active Parameters17.0B per token
ArchitectureMixture of Experts
Total Experts16
Active Experts2 per token
Attention TypeGQA (Grouped Query Attention)
Hidden Dimensiond = 5,120
Transformer Layers48
Attention Heads40
KV Headsn_kv = 8
Head Dimensiond_head = 128
Activation FunctionSwiGLU
NormalizationRMSNorm
Position EmbeddingRoPE

System Requirements

Estimated VRAM at 10% overhead for different quantization methods and context sizes.

Quantization1K ctx195K ctx256K ctx
Q4_K_M0.50 B/W
~97% of FP16
56.53Datacenter GPU
92.96Cluster / Multi-GPU
104.3Cluster / Multi-GPU
Q8_01.00 B/W
~100% of FP16
112.9Cluster / Multi-GPU
149.3Cluster / Multi-GPU
160.7Cluster / Multi-GPU
F162.00 B/W
Reference
225.5Cluster / Multi-GPU
262.0Cluster / Multi-GPU
273.4Cluster / Multi-GPU
Fits 24 GB consumer GPU
Fits 80 GB datacenter GPU
Requires cluster / multi-GPU

Other Llama Models

View All

Find the right GPU for Llama 4 Scout (MoE)

Use the interactive VRAM Calculator to see exactly how much memory you need at any quantization level, context length, and overhead setting.