Table of Contents
Vision-language models accept images as input alongside text. Strongest open-weight options in 2026: Llama 3.2 Vision (11B/90B), Qwen 2.5 VL (3B/7B/72B), MiniCPM-V 2.6 (8B). Each has hardware and licensing nuances.
For most VLM workloads: Qwen 2.5 VL 7B at FP8 on RTX 5090. For Llama-family: Llama 3.2 11B Vision FP16 on 5090. For frontier 90B vision: 6000 Pro or multi-GPU.
Vision-language model landscape
- Qwen 2.5 VL 7B — strongest 7B-class VLM. Document analysis, OCR, charts.
- Qwen 2.5 VL 72B — frontier open VLM.
- Llama 3.2 11B Vision — Meta's vision variant.
- Llama 3.2 90B Vision — frontier Meta vision.
- MiniCPM-V 2.6 (8B) — efficient, edge-friendly.
- Pixtral 12B — Mistral's VLM.
Hardware sizing
| Model | VRAM (FP8) | Recommended GPU |
|---|---|---|
| Qwen 2.5 VL 3B | ~5 GB | RTX 5060 Ti 16 GB |
| Qwen 2.5 VL 7B | ~9 GB | RTX 5060 Ti or 5080 |
| Llama 3.2 11B Vision | ~14 GB | RTX 5080 or 5090 |
| Pixtral 12B | ~14 GB | RTX 5080 or 5090 |
| Qwen 2.5 VL 72B | ~72 GB | RTX 6000 Pro |
| Llama 3.2 90B Vision | ~90 GB | RTX 6000 Pro (tight) or multi-GPU |
Setup with vLLM
vllm serve Qwen/Qwen2.5-VL-7B-Instruct \
--quantization fp8 \
--max-model-len 32768 \
--gpu-memory-utilization 0.92 \
--limit-mm-per-prompt image=4 \
--served-model-name qwen-vlClient uses OpenAI vision message format with {"type":"image_url","image_url":...}.
Verdict
For most VLM use cases (document Q&A, image analysis, chart reading), Qwen 2.5 VL 7B on a 5090 is the sweet spot. Llama 3.2 Vision is fine if you need Meta-family compatibility.
Bottom line
Open-weight VLMs are now genuinely competitive with closed providers for most workloads. See Qwen-VL benchmark on 5060 Ti.
