Trinity Mini
12B
- Context
- 128Ktokens
- Released
- 2y ago
- Size
- 90Bparameters
Llama-3.2-90B-Vision-Instruct is an instruction-tuned vision model from Meta, part of the Llama 3.2 collection that the company framed as the follow-up to Llama 3.1. Meta's announcement headlined open, customizable models for edge AI and vision, saying the vision models are drop-in replacements for their text equivalents while, by its claim, exceeding closed models like Claude 3 Haiku on image understanding. The 90B model is built for image reasoning tasks such as understanding charts and graphs at the document level, captioning images, and visual grounding.
- Multimodal model
- Open weights · 90B
- Accepts images
- Tools
How it compares
Size class
Medium90BAn open-weight model of 25 to 100 billion parameters, for a workstation or a single server GPU.
Open models of a similar size from other labs
Vs. your stack
Add the models you use to see how this one compares.
Sign inAdditional source evidence
- Source model
- meta/llama-3.2-90b-vision-instruct
- Source family
- llama
- Last updated
- 2024-09-25
- Knowledge cutoff
- 2023-12
- Model context
- 128K
- Model output limit
- 4K
- Open weights
- Yes
- File attachments
- Yes
- Reasoning
- No
- Tool calling
- Yes
- Temperature control
- Yes
Downloadable weights
Complete source record (JSON)
{
"id": "meta/llama-3.2-90b-vision-instruct",
"name": "Llama-3.2-90B-Vision-Instruct",
"description": "Large open multimodal Llama model for image reasoning, document QA, and captioning",
"family": "llama",
"last_updated": "2024-09-25",
"knowledge": "2023-12",
"attachment": true,
"reasoning": false,
"tool_call": true,
"temperature": true,
"modalities": {
"input": [
"text",
"image"
],
"output": [
"text"
]
},
"limit": {
"context": 128000,
"output": 4096
},
"release_date": "2024-09-25",
"open_weights": true,
"license": "Llama 3.2 Community License",
"weights": [
{
"label": "Hugging Face",
"url": "https://huggingface.co/meta-llama/Llama-3.2-90B-Vision-Instruct"
}
]
}Frequently asked questions
Running locally
By our estimate, Llama-3.2-90B-Vision-Instruct needs about 63.2 GB at Q4, 110.5 GB at Q8 and 207.5 GB at FP16 of VRAM. That covers the weights plus headroom for a short context, and longer contexts need more.
No, by our estimates Llama-3.2-90B-Vision-Instruct does not fit fully on a 24 GB GPU even at Q4. It needs more VRAM or offloading to system memory.
Yes, a Mac with at least 96 GB of unified memory should run Llama-3.2-90B-Vision-Instruct at Q4, which needs an estimated 63.2 GB. This assumes macOS gives the GPU about 75% of unified memory.
Privacy and data
Yes, Llama-3.2-90B-Vision-Instruct has open weights, so you can run it on your own hardware and keep prompts on your own machines.
Capabilities and specs
Llama-3.2-90B-Vision-Instruct has a 128K token context window. It can write up to 4K output tokens per response.
Yes, Llama-3.2-90B-Vision-Instruct accepts images as input alongside text.
Yes, Llama-3.2-90B-Vision-Instruct supports tool calling (function calling), so it can request calls to tools you define.
No, Llama-3.2-90B-Vision-Instruct is not listed with a reasoning mode.
Llama-3.2-90B-Vision-Instruct has a knowledge cutoff of 2023-12, so it does not know about later events unless you provide them.
Quality and benchmarks
Llama-3.2-90B-Vision-Instruct ranks #148 on Arena Vision, using Arena's published ranks.
Similar models from other labs include Molmo 72B (Ai2), Molmo 7B-D (Ai2) and Qwen2.5 7B Instruct (Alibaba). They are the closest matches by kind and release date.
Molmo 72BLlama-3.2-90B-Vision-Instruct vs Molmo 72BMolmo 7B-DLlama-3.2-90B-Vision-Instruct vs Molmo 7B-DQwen2.5 7B InstructLlama-3.2-90B-Vision-Instruct vs Qwen2.5 7B Instruct
Versions and lineage
Llama-3.2-90B-Vision-Instruct has a 128K token context window, versus 60K for Llama 3.2 1B Instruct.
Llama 3.2 1B InstructLlama-3.2-90B-Vision-Instruct vs Llama 3.2 1B Instruct
Licensing and access
Llama-3.2-90B-Vision-Instruct has open weights under the Llama 3.2 Community License license, so you can download and run it yourself.
The Llama-3.2-90B-Vision-Instruct weights are on Hugging Face at meta-llama/Llama-3.2-90B-Vision-Instruct.
Basics
Llama-3.2-90B-Vision-Instruct was made by Meta. It was released on September 25, 2024.
Yes, Llama-3.2-90B-Vision-Instruct has open weights, so you can run it on your own hardware without API fees.
Generated from Vertical Tabs data, updated November 10, 2026.
Discussion
No comments yet
Be the first to share your thoughts.