Skip to content
Updated minutes ago
Meta

Llama 4 Maverick

Meta · moe · 400B parameters · 1,048,576 context

Quality
84.0

Parameters

400B

Context Window

1024K tokens

Architecture

MoE

Best GPU

B200 SXM

Cheapest API

$1.80/M

Quality Score

84/100

Intelligence Brief

Llama 4 Maverick is a 400B parameter Mixture-of-Experts (128 experts, 1 active) model from Meta, featuring Grouped Query Attention (GQA) with 96 layers and 5,120 hidden dimensions. With a 1,048,576 token context window, it supports tools, vision, structured output, code, math, multilingual. On standardized benchmarks, it achieves MMLU 89, HumanEval 63, GSM8K 95. The most cost-effective API deployment is via together at $1.80/M output tokens. For self-hosted inference, B200 SXM delivers optimal throughput at $17044/month.

Architecture Details

TypeMOE
Total Parameters400B
Active Parameters17B
Layers96
Hidden Dimension5,120
Attention Heads40
KV Heads8
Head Dimension128
Vocab Size202,048
Total Experts128
Active Experts1

Memory Requirements

BF16 Weights

800.0 GB

FP8 Weights

400.0 GB

INT4 Weights

200.0 GB

KV-Cache per Token393216 bytes
Activation Estimate3.00 GB

Fits on (multi-GPU with Tensor Parallelism)

Multi-GPU configurations use Tensor Parallelism (TP) to split model layers across GPUs. Requires NVLink or NVSwitch interconnect for optimal performance.

GPU Compatibility Matrix

Llama 4 Maverick is compatible with 2% of GPU configurations across 41 GPUs at 3 precision levels.

BF16 (Full)
FP8 (Half)
INT4 (Quarter)
Blackwell(7 GPUs)
B200 NVL (pair)360GB
B300288GB
B100 SXM192GB
GB200 NVL72 (per GPU)192GB
Hopper(7 GPUs)
H100 NVL 94GB (per GPU pair)188GB
H200 SXM141GB
H2096GB
GH20096GB
Ada Lovelace(11 GPUs)
L40S48GB
L4048GB
RTX 6000 Ada48GB
L2048GB
Ampere(16 GPUs)
A100 80GB SXM80GB
A100 80GB PCIe80GB
A1664GB
RTX A600048GB
Legend:No fitVery tightTightModerateGoodExcellent

GPU Recommendations

B200 SXMoptimal

FP8 · 4 GPUs · tensorrt-llm

100/100

score

Throughput

280.0 tok/s

Latency (ITL)

3.6ms

Est. TTFT

1ms

Cost/Month

$17044

Cost/M Tokens

$23.16

Use this config →
B100 SXMoptimal

FP8 · 4 GPUs · tensorrt-llm

100/100

score

Throughput

280.0 tok/s

Latency (ITL)

3.6ms

Est. TTFT

1ms

Cost/Month

$17082

Cost/M Tokens

$23.21

Use this config →
H200 SXMoptimal

FP8 · 4 GPUs · tensorrt-llm

100/100

score

Throughput

280.0 tok/s

Latency (ITL)

3.6ms

Est. TTFT

1ms

Cost/Month

$10211

Cost/M Tokens

$13.88

Use this config →

Deployment Options

API

API Deployment

together

$1.80/M

output tokens

Self-Hosted

Single GPU

Requires multi-GPU setup (400 GB VRAM needed)

Scale

Multi-GPU

B200 SXM x4

280.0 tok/s

TP· $17044/mo

API Pricing Comparison

ProviderInput $/MOutput $/MBadges
together$1.20$1.80
Cheapest
fireworks$1.50$2.00

Cost Analysis

ProviderInput $/MOutput $/M~Monthly Cost
togetherBest Value$1.20$1.80$15
fireworks$1.50$2.00$18

Cost per 1,000 Requests

Short (500 tok)

$0.96

via together

Medium (2K tok)

$3.84

via together

Long (8K tok)

$13.20

via together

Performance Estimates

Throughput by GPU

B200 SXM
280.0 tok/s
B100 SXM
280.0 tok/s
H200 SXM
280.0 tok/s

VRAM Breakdown (B200 SXM, FP8)

Weights
Act
Weights 100.0 GBKV-Cache 3.2 GBActivations 24.0 GBOverhead 5.0 GB

Precision Impact

bf16

200.0 GB

weights/GPU

fp8

100.0 GB

weights/GPU

~280.0 tok/s

int4

50.0 GB

weights/GPU

Quality Benchmarks

Top 10%
93th percentile across all models
MMLU
89.0
Above Average (89th pctile)
HumanEval
63.0
Above Average (75th pctile)
GSM8K
95.0
Above Average (87th pctile)
MT-Bench
88.0
Bottom 25% (0th pctile)

Capabilities

Features

Tool Use Vision Code Math Reasoning Multilingual Structured Output

Supported Frameworks

vllmsglangtensorrt-llm

Supported Precisions

BF16 (default)FP8INT4

Where to Deploy Llama 4 Maverick

Similar Models

Smaller context, Lower quality, More expensiveCompare →

Llama 3.1 405B

405B params · dense

Quality: 81

from $3.00/M

Smaller context, More expensiveCompare →
Smaller context, Lower qualityCompare →

MiniMax-Text-01

456B params · moe

Quality: 50

from $5.00/M

Lower quality, More expensiveCompare →

MiniMax M2.7

456B params · moe

Quality: 82

from $2.80/M

More expensiveCompare →

Frequently Asked Questions

How much VRAM does Llama 4 Maverick need for inference?

Llama 4 Maverick requires approximately 800.0 GB of VRAM at BF16 precision, 400.0 GB at FP8, or 200.0 GB at INT4 quantization. Additional VRAM is needed for KV-cache (393216 bytes per token) and activations (~3.00 GB).

What is the best GPU for Llama 4 Maverick?

The top recommended GPU for Llama 4 Maverick is the B200 SXM (x4) using FP8 precision. It achieves approximately 280.0 tokens/sec at an estimated cost of $17044/month ($23.16/M tokens). Score: 100/100.

How much does Llama 4 Maverick inference cost?

Llama 4 Maverick API inference starts from $1.20/M input tokens and $1.80/M output tokens. Self-hosted inference costs depend on your GPU configuration — use our ROI calculator for a detailed breakdown.