Model

Llama 4 Scout

OPEN

Meta · Llama 4 family · released Apr 5, 2025

Meta's first MoE Llama; 17B active over 109B total (16 experts) with an industry-leading 10M-token context window.

ReasoningCodingVisionFunction callingTool useAgentic

1054.1

Elo · rank #203

Parameters

109B

Active params

17B (MoE)

Context

10M tokens

Architecture

MoE, 16 experts with early-fusion native multimodality, 17B active / 109B total, 10M-token context

License

Llama 4 Community License Agreement

Languages

12+

API price (in/out)

No hosted API

Modalities

text · vision

Benchmark results

Bar shows position within the tracked field; marker = field best

ChartQAVision88.8%#8

best: MiniMax-VL-01 · 91.7%

DocVQAVision94.4%#10

best: Qwen2-VL-72B · 96.5%

GPQA DiamondReasoning57.2%#169

best: GPT-5.6 · 94.6%

LiveCodeBenchCoding32.8%#144

best: DeepSeek-V4-Pro (Think Max) · 93.5%

MATH-500Math50.3%#132

best: GPT-5 · 99.4%

MathVistaVision70.7%#25

best: Seed 2.1 Pro · 90.7%

MBPPCoding67.8%#43

best: Llama-3.3-Nemotron-Super-49B v1 (Reasoning On) · 91.3%

MGSMMath90.6%#11

best: OpenAI o4-mini · 93.7%

MMLU-ProKnowledge74.3%#77

best: Claude Fable 5 · 91.5%

MMLUKnowledge79.6%#94

best: OpenAI o3 · 92.9%

MMMU-ProVision52.2%#43

best: Claude Opus 4.7 · 85.5%

MMMUVision69.4%#57

best: Claude Fable 5 · 89.3%

Run it locally

VRAM @ Q4

58 GB

VRAM @ FP16

218 GB

Fits on (Q4)

M3 Max 128GBM3 Ultra 512GBA100 80GBH100 80GBH200 141GBB200 192GB

Fits a single H100 80GB at Q4 (~58GB) but still exceeds a 24GB RTX 4090; no single-4090 llama.cpp benchmark exists.

Quantizations

GGUF · AWQ · MLX

Llama 4 family

Elo progression across releases

Llama 4 MaverickApr 20251253.3 Llama 4 ScoutApr 20251054.1

Resources

HFmeta-llama/Llama-4-Scout-17B-16E-Instruct GHmeta-llama/llama-models DOCSllama.com/models/llama-4 APIAvailable via Meta API

API price weights · each benchmark row carries its own source badge (see methodology)