Skip to content
Updated minutes ago
Meta

Llama 3.2 90B Vision Instruct

Meta · dense · 88.8B parameters · 131,072 context

Quality
84.0

Parameters

88.8B

Context Window

128K tokens

Architecture

Dense

Best GPU

B200 SXM

Cheapest API

$1.20/M

Quality Score

84/100

Intelligence Brief

Llama 3.2 90B Vision Instruct is a 88.8B parameter DENSE model from Meta, featuring Grouped Query Attention (GQA) with 80 layers and 8,192 hidden dimensions. With a 131,072 token context window, it supports tools, vision, structured output, code, math, multilingual. On standardized benchmarks, it achieves MMLU 86, HumanEval 58, GSM8K 92. The most cost-effective API deployment is via together at $1.20/M output tokens. For self-hosted inference, B200 SXM delivers optimal throughput at $4261/month.

Architecture Details

TypeDENSE
Total Parameters88.8B
Active Parameters88.8B
Layers80
Hidden Dimension8,192
Attention Heads64
KV Heads8
Head Dimension128
Vocab Size128,256

Memory Requirements

BF16 Weights

177.6 GB

FP8 Weights

88.8 GB

INT4 Weights

44.4 GB

KV-Cache per Token655360 bytes
Activation Estimate4.00 GB

GPU Compatibility Matrix

Llama 3.2 90B Vision Instruct is compatible with 33% 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 · 1 GPU · tensorrt-llm

100/100

score

Throughput

560.0 tok/s

Latency (ITL)

1.8ms

Est. TTFT

0ms

Cost/Month

$4261

Cost/M Tokens

$2.90

Use this config →
H200 SXMoptimal

FP8 · 1 GPU · tensorrt-llm

100/100

score

Throughput

452.2 tok/s

Latency (ITL)

2.2ms

Est. TTFT

0ms

Cost/Month

$2553

Cost/M Tokens

$2.15

Use this config →
H100 SXMoptimal

FP8 · 2 GPUs · tensorrt-llm

100/100

score

Throughput

560.0 tok/s

Latency (ITL)

1.8ms

Est. TTFT

0ms

Cost/Month

$3587

Cost/M Tokens

$2.44

Use this config →

Deployment Options

API

API Deployment

together

$1.20/M

output tokens

Self-Hosted

Single GPU

B200 SXM

$4261/mo

Min VRAM: 89 GB

Scale

Multi-GPU

H100 SXM x2

560.0 tok/s

TP· $3587/mo

API Pricing Comparison

ProviderInput $/MOutput $/MBadges
together$1.20$1.20
Cheapest

Cost Analysis

ProviderInput $/MOutput $/M~Monthly Cost
togetherBest Value$1.20$1.20$12

Cost per 1,000 Requests

Short (500 tok)

$0.84

via together

Medium (2K tok)

$3.36

via together

Long (8K tok)

$12.00

via together

Performance Estimates

Throughput by GPU

B200 SXM
560.0 tok/s
H200 SXM
452.2 tok/s
H100 SXM
560.0 tok/s

VRAM Breakdown (B200 SXM, FP8)

Weights
Act
Weights 88.8 GBKV-Cache 2.7 GBActivations 32.0 GBOverhead 4.4 GB

Precision Impact

bf16

177.6 GB

weights/GPU

fp8

88.8 GB

weights/GPU

~560.0 tok/s

int4

44.4 GB

weights/GPU

Quality Benchmarks

Top 10%
93th percentile across all models
MMLU
86.0
Above Average (78th pctile)
HumanEval
58.0
Average (67th pctile)
GSM8K
92.0
Above Average (76th pctile)
MT-Bench
86.0
Bottom 25% (0th pctile)

Capabilities

Features

Tool Use Vision Code Math Reasoning Multilingual Structured Output

Supported Frameworks

vllmsglangtgitensorrt-llm

Supported Precisions

BF16 (default)FP8INT4

Where to Deploy Llama 3.2 90B Vision Instruct

Similar Models

Similar specsCompare →

Inflection 3

100B params · dense

Quality: 74

from $15.00/M

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

Yi-Large

102.6B params · moe

Quality: 74

from $3.00/M

Smaller context, Lower quality, More expensiveCompare →

Command R+

104B params · dense

Quality: 68

from $2.00/M

Lower quality, More expensiveCompare →

Frequently Asked Questions

How much VRAM does Llama 3.2 90B Vision Instruct need for inference?

Llama 3.2 90B Vision Instruct requires approximately 177.6 GB of VRAM at BF16 precision, 88.8 GB at FP8, or 44.4 GB at INT4 quantization. Additional VRAM is needed for KV-cache (655360 bytes per token) and activations (~4.00 GB).

What is the best GPU for Llama 3.2 90B Vision Instruct?

The top recommended GPU for Llama 3.2 90B Vision Instruct is the B200 SXM using FP8 precision. It achieves approximately 560.0 tokens/sec at an estimated cost of $4261/month ($2.90/M tokens). Score: 100/100.

How much does Llama 3.2 90B Vision Instruct inference cost?

Llama 3.2 90B Vision Instruct API inference starts from $1.20/M input tokens and $1.20/M output tokens. Self-hosted inference costs depend on your GPU configuration — use our ROI calculator for a detailed breakdown.