Table of Contents
Llama 3.2 11B Vision is Meta's smallest vision-language model. ~22 GB at FP16 (does not fit 16 GB), ~11 GB at FP8 (fits comfortably), ~6 GB at AWQ-INT4 (lots of room).
TL;DR
Llama 3.2 11B Vision at FP8 fits the RTX 5060 Ti with ~5 GB headroom. ~3.2 image-Q&A queries/sec single-stream. For higher concurrency or 90B Vision, step up to a 5090 or 6000 Pro.
VRAM fit
| Precision | VRAM (weights) | Fits 5060 Ti? |
|---|---|---|
| FP16 | 22 GB | No |
| FP8 | 11 GB | Yes |
| AWQ-INT4 | 6 GB | Yes |
Performance
| Workload | 5060 Ti FP8 |
|---|---|
| Image Q&A (1024×1024 + prompt) | ~3.2/sec |
| OCR (A4 page) | ~5 sec/page |
| Concurrent Q&A throughput (8 streams) | ~12/sec aggregate |
Verdict
Llama 3.2 11B Vision on a 5060 Ti is workable for low-concurrency vision-language workloads. For most VLM use cases Qwen 2.5 VL 7B is a stronger, smaller alternative — see Qwen-VL benchmark.
Bottom line
For Llama-family compatibility, 11B Vision at FP8 on the 5060 Ti works. For pure quality, Qwen 2.5 VL 7B is competitive at less VRAM.
