Model

Llama 4 Maverick

OPEN

Meta · Llama 4 family · released Apr 5, 2025

Natively multimodal MoE flagship; 17B active over 400B total across 128 experts, 1M-token context.

ReasoningCodingVisionFunction callingTool useAgentic

1216.1

Elo · rank #187

Parameters

400B

Active params

17B (MoE)

Context

1M tokens

Architecture

MoE, 128 experts (routed + shared) with early-fusion native multimodality, 17B active / 400B total

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

ChartQAVision85.3%#24

best: MiniMax-VL-01 · 91.7%

DocVQAVision91.6%#28

best: Qwen2-VL-72B · 96.5%

GPQA DiamondReasoning69.8%#132

best: GPT-6 Astra · 96.0%

LiveCodeBenchCoding43.4%#130

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

MATH-500Math61.2%#110

best: GPT-5 · 99.4%

MathVistaVision73.7%#18

best: Seed 2.1 Pro · 90.7%

MBPPCoding77.6%#21

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

MGSMMath92.3%#3

best: OpenAI o4-mini · 93.7%

MMLU-ProKnowledge80.5%#55

best: Claude Fable 5 · 91.5%

MMLUKnowledge85.5%#55

best: OpenAI o3 · 92.9%

MMMU-ProVision59.6%#44

best: Claude Opus 4.7 · 85.5%

MMMUVision73.4%#40

best: Claude Fable 5 · 89.3%

Run it locally

VRAM @ Q4

210 GB

VRAM @ FP16

800 GB

Fits on (Q4)

M3 Ultra 512GB

400B total MoE weights (~210GB at Q4) far exceed a single RTX 4090; needs multi-GPU/datacenter hardware — no single-4090 benchmark exists.

Quantizations

GGUF · AWQ · MLX

Fine-tune it

Conditional / custom

QLoRA272.0 GB2× H200 141GB

LoRA852.0 GB8× H200 141GB

Full fine-tune6420.0 GBbeyond 8× B200

QLoRA SFT on ~10k samples ≈ $8.79 (2× H200 141GB)

Plan a fine-tune →

Llama 4 family

Elo progression across releases

Llama 4 MaverickApr 20251216.1 Llama 4 ScoutApr 2025984.3

Resources

HFmeta-llama/Llama-4-Maverick-17B-128E-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)