3 min read - Llama 3.2 Vision: A Deployment and Evaluation Guide

Meta released Llama 3.2 in September 2024. Its announcement describes vision-capable 11B and 90B models for tasks such as document understanding, image captioning and visual grounding. Those are release facts, not proof that either model will meet a production requirement.

Earlier versions of this article called the release the first open model to rival specific proprietary systems and published fixed VRAM requirements. Those claims were too broad and have been removed.

Use the correct licence description

Meta makes the weights available under its Llama licence and acceptable-use policy. “Open weights” is a safer description than assuming the release meets every definition of open-source software. Review the current licence for the model and intended use, including redistribution and deployment conditions.

Downloading weights also creates a provenance duty. Record the official repository, revision, file hashes and any conversion or quantisation applied before deployment.

Start with one visual task

Image understanding covers different problems. A chart question, scanned invoice, product photo and screenshot do not share the same error profile. Choose one task and define:

• accepted image types and quality;
• the required output schema;
• what counts as a correct answer;
• sensitive information that must be removed or protected;
• the human review and escalation path;
• latency and resource limits on target hardware.

Build an evaluation set from representative, authorised images. Include blur, rotation, small text, unusual layouts and cases where the correct response is to abstain.

Measure deployment rather than copying a hardware table

Memory depends on weight format, runtime, context, image processing, cache, batch size and concurrency. Test the exact artefact on the target environment. Record peak memory, startup time, time to first output, throughput under expected concurrency and failure behaviour near resource limits.

Quantisation can reduce model size, but it can also affect task quality. Compare each candidate format against the same visual evaluation set. A generic percentage from another model or benchmark is not a capacity plan.

Treat image input as untrusted

Images and extracted text may contain sensitive data or instructions intended to influence a downstream agent. Decode files in an isolated component, enforce size and type limits, remove unnecessary metadata and never grant tools solely because the model inferred an instruction from an image.

If the output drives a business action, validate it against an explicit schema and policy. Preserve the source image identifier, model revision, prompt, output and reviewer decision in the audit record.

Compare with the actual alternative

Do not compare Llama 3.2 Vision with a remembered snapshot of another model. Run the same authorised test set against the deployment options available to the team on the review date. Score task accuracy, abstention, latency, operating cost, data path and licence constraints.

The right result may be a local open-weight model, a hosted API, a specialised OCR pipeline or a combination. The model name does not decide the architecture.

Primary source

• Meta's Llama 3.2 release announcement, reviewed 27 August 2026