Model

Llama 3.2 90B Vision

OPEN

Meta · Llama 3.2 family · released Sep 25, 2024

Multimodal 90B-class vision flagship (128K ctx) rivaling closed models like Claude 3 Haiku on image reasoning tasks.

ReasoningCodingVisionFunction callingTool useAgentic

999.5

Elo · rank #213

Parameters

88.8B

Active params

88.8B (dense)

Context

128K tokens

Architecture

Dense Llama 3.1 text backbone + cross-attention vision adapter (late fusion)

License

Llama 3.2 Community License Agreement

Languages

API price (in/out)

No hosted API

Modalities

text · vision

Benchmark results

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

AI2DVision92.3%#17

best: Molmo 72B · 96.3%

ChartQAVision85.5%#22

best: MiniMax-VL-01 · 91.7%

DocVQAVision90.1%#32

best: Qwen2-VL-72B · 96.5%

GPQA DiamondReasoning46.7%#200

best: GPT-5.6 · 94.6%

MATH-500Math68.0%#101

best: GPT-5 · 99.4%

MathVistaVision57.3%#47

best: Seed 2.1 Pro · 90.7%

MGSMMath86.9%#21

best: OpenAI o4-mini · 93.7%

MMLUKnowledge86.0%#46

best: OpenAI o3 · 92.9%

MMMU-ProVision33.8%#50

best: Claude Opus 4.7 · 85.5%

MMMUVision60.3%#77

best: Claude Fable 5 · 89.3%

VQAv2Vision78.1%#12

best: Reka Edge (2026) · 88.4%

Run it locally

VRAM @ Q4

55 GB

VRAM @ FP16

178 GB

Fits on (Q4)

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

Even at Q4 (~55GB) it exceeds a single RTX 4090's 24GB VRAM; no verified single-4090 llama.cpp benchmark exists — needs multi-GPU or CPU offload.

Quantizations

GGUF · AWQ · GPTQ · MLX

Llama 3.2 family

Elo progression across releases

Llama 3.2 11B VisionSep 2024682.7 Llama 3.2 1BSep 202481.3 Llama 3.2 3BSep 2024380.2 Llama 3.2 90B VisionSep 2024999.5

Resources

HFmeta-llama/Llama-3.2-90B-Vision-Instruct GHmeta-llama/llama-models DOCSllama.com/docs/model-cards-and-prompt-formats/llama3_2 APIAvailable via Meta API

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