Llama 3.2 90B Vision Instruct

Llama 3.2 90B Vision Instruct is a multimodal instruction-tuned model developed by Meta. It combines the Llama 3.1 text model with a vision adapter for image reasoning and supports both image and text input. It is part of the Llama 3.2-Vision model collection.

This model excels in:

• Visual Question Answering (VQA)
• Chart, diagram, and infographic analysis
• Document understanding (DocVQA)
• Image captioning and image-text retrieval
• Visual grounding and reasoning tasks
• Agentic systems that process image + text input

It supports a maximum context length of 131,072 tokens on Fireworks.

Fireworks supports the full 131K token window on dedicated GPU deployments (on-demand).

• No support for structured function calling or tool use
• May hallucinate or show reduced precision in adversarial visual prompts
• Supports only English for image-text tasks
• Ethical risks include child safety, cyber attacks, and content misuse; mitigated with fine-tuning and Llama Guard

The model has 88.6 billion parameters.

Fine-tuning is supported via LoRA on Fireworks' serverless interface, but standard fine-tuning is not available.

Tokens are counted across combined input and output, up to the 131K limit.

• Serverless: Not supported
• On-demand: Available with no rate limits on dedicated infrastructure

The model is governed by the Llama 3.2 Community License, a custom commercial license that allows research and commercial use under specified terms.