Meta logo Open weights · Llama 3.2 Community License

Llama 3.2 90B Vision

The largest multimodal Llama 3.2 at 90B parameters, adding image reasoning through cross-attention adapters on the Llama 3.1 70B backbone.

Jump to: Specs Pricing Self-hosting FAQ More from Meta
Visit website Compare vs other LLMs →

Key Specifications

Context window
128K tokens
Knowledge cutoff
Parameters
88.8B
Inputs
Text, image
Capabilities Show details
Function calling Connect to external tools, APIs, and systems.

Hosted API pricing

Provider Input / 1M tokens Output / 1M tokens Cost at 10M in + 2M out
Meta logo Meta Creator Open weights, no hosted price View

Heads up: Base-tier, on-demand rates per 1M tokens; cached, batch and long-context tiers excluded. A provider may serve a shorter context or a quantized build than the creator's release. Verify before provisioning. More on how we price.

Estimated cost to self-host

Llama 3.2 90B Vision needs about 56 GB of GPU memory at 4-bit with 32K context. The cheapest rental that fits is 4× RTX 4060 Ti at about $317 a month.

Single node ≤ 8
Precision Memory Cheapest fit (32K context) Cost / mo
4-bitINT4 / FP4
56 GB
Nvidia logo 4× RTX 4060 Ti
$317
8-bitFP8 / INT8
96 GB
Nvidia logo 8× RTX 4060 Ti
$634
16-bitFP16 / BF16
190 GB
Nvidia logo 8× RTX A6000
$3,168

Estimates based on median on-demand rates for Nvidia GPUs. Memory is weights plus KV cache for one request, using FP8 KV cache where supported and FP16/BF16 otherwise. No guarantee of runtime support, usable performance, or that a matching quantized build exists. How we estimate costs.

Thunder Compute Logo Thunder Compute
Our sponsor
RTX A6000 from $0.35, A100 from $1.09, per GPU per hour
Billed per minute, no commitment, no minimums
One click into VS Code or Cursor, no SSH setup
See live pricing

Frequently Asked Questions

What is Llama 3.2 90B Vision good for?

Image reasoning with text, at the heavy end of the Llama 3.2 family. Open weights under the Llama 3.2 Community License, servable on a single high-memory GPU.

When is Llama 3.2 90B Vision not a good fit?

Tool calling on image inputs: it works only with text-only prompts. Its knowledge is older than the current open models', and self-hosting takes a high-memory GPU rather than a consumer card.

What is the cheapest way to run Llama 3.2 90B Vision?

No provider we track hosts it; self-hosting on the cheapest rental that fits (4x RTX 4060 Ti, $0.44 per hour) costs about $317 a month.

Can I self-host Llama 3.2 90B Vision?

Yes. The weights are released under the Llama 3.2 Community License. At 4-bit it needs about 56 GB of GPU memory, which starts at roughly $317 a month on the cheapest rental that fits.

More from Meta

Model Context Input / 1M Output / 1M
Meta Muse Glimmer 30B Open weights 131K $0.20 $0.80
Meta Llama 4 Scout Open weights 10M $0.16 $0.64
Meta Llama 4 Maverick Open weights 1M $0.20 $0.70
Meta Llama 3.1 8B Open weights 131K $0.02 $0.05
Meta Llama 3.2 11B Vision Open weights 128K $0.05 $0.68
Meta Llama 3.2 1B Open weights 128K $0.03 $0.20
Meta Llama 3.2 3B Open weights 128K $0.05 $0.34
Meta Llama 3.3 70B Open weights 128K $0.13 $0.40