Llama 3.2 90B Vision Instruct

Positioning

Llama 3.2 90B Vision Instruct is Meta's multimodal vision-language model — the instruction-tuned variant of the 90B Vision family. As Meta's largest open-weight VL model, it combines the Llama 3.2 90B text backbone with a vision encoder capable of understanding images, charts, documents, and photographs. For operators who need visual understanding at scale with open weights and a permissive-enough license (subject to the Llama 3.2 Community License), this is the reference model in the 70-100B VL class.

Strengths

• Largest open-weight vision-language model from Meta. 90B parameter text backbone with a dedicated vision encoder provides strong visual understanding across diverse image types — photographs, screenshots, documents, diagrams, and charts.
• Instruction-tuned for vision tasks. Pre-tuned for visual Q&A, image description, document understanding, and chart analysis. Usable out of the box for visual tasks without additional fine-tuning.
• Strong text reasoning capability from Llama 3.2 backbone. The text-only reasoning (MMLU-Pro, HumanEval, GSM8K) is competitive with pure text models at the 70B scale — the vision capability is additive, not a tradeoff.
• Llama ecosystem compatibility. Works with vLLM, SGLang, Ollama, and llama.cpp with vision support. Broad tooling ecosystem makes deployment straightforward for Llama-ecosystem operators.
• 90B fits workstation hardware. Q4 at ~52 GB fits a single A100 80GB or Mac Studio M3 Ultra. Q3 at ~40 GB fits dual RTX 4090 setups. More accessible than the 123B+ VL models.

Limitations

• Llama 3.2 Community License restrictions. The 700M MAU clause and use restrictions apply. For Apache 2.0 VL alternatives, consider InternVL 2.5 78B (verify license).
• Vision encoder resolution limits. The vision encoder's native resolution determines the maximum image detail the model can perceive. For high-resolution document OCR or detailed medical image analysis, dedicated vision models may be more appropriate than a general-purpose VL model.
• Video understanding is limited. The model processes images (single frames), not video. For video understanding, consider models with temporal awareness or process individual frames with the model's image capability.
• Not a dedicated OCR or document model. While capable of document understanding, specialized models like Qwen 2.5-VL 72B or dedicated OCR pipelines may provide higher accuracy for document-heavy workflows.
• Hardware requirements including vision encoder. The vision encoder adds memory overhead beyond the text model. Total VRAM requirements are slightly higher than the 90B text-only model — budget 5-10 GB additional for vision processing at higher resolutions.

Real-world performance

• vs Llama 3.2 90B Vision (base): Instruct vs base. The instruct variant is chat-ready for vision tasks; the base variant is for fine-tuning. Pick Instruct for deployment; pick Base for custom VL training pipelines.
• vs InternVL 2.5 78B: Both open-weight VL models in the 70-100B range. InternVL is purpose-built for vision-language with a stronger vision encoder. Llama 3.2 90B Vision has stronger text reasoning from the Llama backbone. Pick by visual emphasis vs text reasoning balance.
• vs Qwen 2.5-VL 72B: Qwen 2.5-VL is the Qwen family's VL flagship with strong Chinese visual understanding. Llama 3.2 90B Vision has stronger English text reasoning and broader ecosystem support.

Should you run this locally?

Yes if you need visual understanding at scale (90B class), the Llama 3.2 Community License works for your deployment, and you have workstation-class hardware (Mac Studio M3 Ultra at Q4, or dual A100). The combination of Llama text reasoning with vision capability is compelling for document Q&A, visual assistants, and image analysis.

No if you need Apache 2.0 licensing (verify InternVL 2.5 78B license), your visual tasks are highly specialized (dedicated vision models may outperform), or you lack workstation-class hardware.

How it compares

• vs Llama 3.2 90B Vision (base): Instruct for deployment; Base for fine-tuning. Same architecture.
• vs InternVL 2.5 78B: InternVL has stronger vision encoding; Llama has stronger text reasoning.
• vs Qwen 2.5-VL 72B: Different language strengths; Qwen stronger on Chinese visual tasks.
• vs Molmo 72B: Molmo adds pointing/grounding capabilities; Llama 90B Vision has larger text backbone.

Run this yourself

• Datacenter: Single A100 80GB at FP8 with vLLM vision support. Production target: 30-50 tok/s with image processing.
• Workstation: Mac Studio M3 Ultra (192 GB) at Q4 — comfortable fit at 10-15 tok/s.
• Dual GPU: 2× RTX 4090 at Q3 with ExLlamaV2 — functional at 15-25 tok/s.
• Vendor: huggingface.co/meta-llama/Llama-3.2-90B-Vision-Instruct