Llama 3.2 Instruct 90B
Llama 3.2 Instruct 90B is Meta's larger Llama 3.2 instruction-tuned model, released with vision capability and offered through multiple hosted inference providers.
API Pricing
| Provider | Input / 1M | Output / 1M |
|---|---|---|
| $0.720 | $0.720 |
Prices updated daily. Last check: Nov 6, 2026
Compare API pricing for every Llama model →Llama 3.2 Instruct 90B pricing by provider
Input, output, and batch rates, plus alternatives, for one provider at a time.
Model Details
General
- Creator
- Meta
- Modalities
- Text
Capabilities
- Open Source
- Yes
Strengths & Limitations
Strengths
- Open-weight release under Meta's Llama community license, allowing self-hosting and on-premise deployment
- Vision-capable: accepts image inputs for tasks like chart, diagram, and document interpretation
- 128K token context window as published by Meta for the Llama 3.2 series
- Largest parameter count in the Llama 3.2 generation, giving more headroom than the 1B, 3B, and 11B variants
- Available from multiple hosted inference providers, so pricing and throughput can be compared rather than fixed to one vendor
- Instruction-tuned for chat and assistant-style use without additional fine-tuning
Limitations
- At roughly 90B parameters it is expensive to self-host, requiring multi-GPU setups with substantial VRAM
- Our database currently records no verified throughput or time-to-first-token measurements for this model
- Superseded in Meta's own roadmap by later Llama generations, so it may not reflect current Llama capabilities
- Higher latency and cost per request than the smaller Llama 3.2 1B, 3B, and 11B variants for simple text tasks
- Behavior and available features (tool calling, JSON mode, image handling limits) vary by hosting provider
Key Features
About Llama 3.2 Instruct 90B
Common Use Cases
Llama 3.2 Instruct 90B suits workloads that combine text and image understanding — extracting information from scanned documents, describing or captioning images, answering questions about charts and screenshots, and multimodal support assistants. Its large context window makes it usable for long-document summarization and retrieval-augmented pipelines where many retrieved passages are stuffed into a single prompt. The open-weight license makes it a common choice for teams that need to run inference inside their own infrastructure for compliance reasons, or that want the option to fine-tune on proprietary data. For high-volume, latency-sensitive text-only classification or routing, the smaller Llama 3.2 3B or 11B variants are generally the more economical fit, with the 90B reserved for the harder reasoning and vision tasks in a tiered setup.
Frequently Asked Questions
How much does Llama 3.2 Instruct 90B cost to use?
Because the weights are openly available, the model is hosted by many different inference providers, and each sets its own rates for input and output tokens; some also offer batch or dedicated-capacity pricing. Rates therefore vary meaningfully between providers. See the pricing table on this page for current per-provider figures, and factor in your own input/output token mix when comparing.
What is Llama 3.2 Instruct 90B best used for?
It fits multimodal tasks that pair text with images — document and chart understanding, image captioning, visual Q&A — plus long-context text work such as summarizing large documents or reasoning over many retrieved passages within its 128K token window. It is also a common pick when open weights and self-hosting are requirements.
How does it differ from the smaller Llama 3.2 models?
The Llama 3.2 release includes small text-only models at 1B and 3B intended for on-device and lightweight use, plus vision-capable models at 11B and 90B. The 90B is the highest-capacity tier of the generation, offering more headroom on complex reasoning at the cost of higher compute, latency, and price than the smaller variants.
Can I run Llama 3.2 Instruct 90B myself?
Yes — Meta distributes Llama 3.2 weights under its community license, so self-hosting and fine-tuning are possible. At roughly 90B parameters this requires substantial GPU memory, typically a multi-GPU server, and quantized builds are often used to reduce hardware requirements. Review the license terms for your specific deployment scenario.
Does this model support image inputs?
The larger Llama 3.2 models, including the 90B, were released with vision capability for image reasoning tasks. However, whether a given hosted endpoint exposes image input — and what resolution or file-size limits apply — depends on the provider, so check the provider's API documentation before building against it.