Llama 3.2 90B Vision
Llama 3.2 multimodal at 90B. Datacenter-tier predecessor to Llama 4 Maverick. Strong visual reasoning.
Our verdict
Positioning
Llama 3.2 90B Vision is Meta's 90B-parameter multimodal base model — the non-instruction-tuned variant of the Llama 3.2 Vision family. Like Llama 3.2 90B Vision Instruct, this model combines a 90B text backbone with a vision encoder capable of understanding images. The key difference: this is the base model, designed for fine-tuning, continued pre-training, and custom visual-language adaptation — not for direct chat or Q&A use.
Strengths
- Largest Llama 3.2 Vision base model from Meta. 90B parameters with vision encoding provides rich capacity for custom visual-language fine-tuning. Operators who need a VL model tailored to their domain (medical imaging, document processing, industrial inspection) have a strong starting point.
- Llama architecture compatibility for fine-tuning. Any fine-tuning pipeline that works with Llama text models can be extended for vision tasks. The broad Llama fine-tuning ecosystem (LoRA, QLoRA, full fine-tune) applies.
- Strong text backbone from Llama 3.2. The 90B text backbone has competitive reasoning and knowledge capability before vision fine-tuning — custom VL training builds on a strong foundation.
- Meta's training infrastructure quality. The base model benefits from Meta's production-scale training pipeline, with well-characterized pre-training data and documented training methodology.
- Llama 3.2 Community License. While not Apache 2.0, the license enables commercial use (subject to MAU clause) and broad redistribution, making it viable for many commercial fine-tuning workflows.
Limitations
- Base model — not usable for vision tasks out of the box. This model requires fine-tuning before it can perform visual Q&A, image description, or any user-facing vision task. Operators looking for ready-to-use VL models should pick the Instruct variant instead.
- Fine-tuning a 90B VL model is expensive. Full fine-tuning requires 4-8× A100 80GB. Even LoRA/QLoRA on VL models at 90B scale is resource-intensive — budget 2-4× A100 for QLoRA fine-tuning.
- Llama 3.2 Community License restrictions. The 700M MAU clause applies. For fully permissive Apache 2.0 VL base models, verify alternatives like InternVL 2.5 78B.
- Vision encoder resolution is fixed. The native vision encoder resolution bounds the maximum image detail the model can learn. For applications requiring very high-resolution image understanding, the base architecture may be a constraint.
- Knowledge cutoff is December 2023. The text backbone's pre-training data doesn't include 2024 or later knowledge, which matters for domain-specific fine-tuning where knowledge recency is important.
Real-world performance
- vs Llama 3.2 90B Vision Instruct: Base vs instruct — same architecture, different training stage. Instruct is ready for deployment; Base is for custom training. The Instruct variant demonstrates what the base model is capable of when fine-tuned properly.
- vs Llama 3.3 70B (text-only, instruction-tuned): Llama 3.2 90B Vision adds vision encoding to a similar-scale text backbone. For text-only fine-tuning, the pure text 70B model is a better starting point. For VL fine-tuning, the 90B Vision base is the right pick.
- vs InternVL 2.5 78B (base): Both VL base models. InternVL's vision encoder is purpose-built and potentially stronger for visual tasks; Llama's text backbone is stronger for reasoning-heavy VL tasks. Pick by the dominant modality in your target task.
Should you run this locally?
Yes if you have a custom visual-language fine-tuning pipeline, need a 90B-class VL base model with Llama ecosystem compatibility, and have the hardware for fine-tuning (4-8× A100 for full fine-tune, 2-4× for QLoRA). The combination of Llama architecture familiarity with vision capability makes this a strong starting point.
No if you want a ready-to-use VL model (pick Llama 3.2 90B Vision Instruct), don't plan to fine-tune, or lack the hardware for 90B VL fine-tuning.
How it compares
- vs Llama 3.2 90B Vision Instruct: Base for fine-tuning; Instruct for direct use. Same architecture.
- vs Llama 3.3 70B: Vision base vs text-only instruct model. Pick by modality need.
- vs InternVL 2.5 78B: Different vision encoder designs. InternVL for visual-first; Llama for text-reasoning-first VL tasks.
Run this yourself
- Fine-tuning: 4-8× A100 80GB for full fine-tune. 2-4× A100 80GB for QLoRA.
- Inference (post fine-tune): Single A100 80GB at Q4 (FP8 weights of about 89 GB exceed its memory), or Mac Studio M3 Ultra at Q4.
- Vendor: huggingface.co/meta-llama/Llama-3.2-90B-Vision
Overview
Llama 3.2 multimodal at 90B. Datacenter-tier predecessor to Llama 4 Maverick. Strong visual reasoning.
How to run it
Llama 3.2 90B Vision is Meta's base multimodal model — 90B dense text backbone with vision encoder, no instruction tuning. Same architecture as the instruct variant but no chat template — generates completions, not responses. Meta's model card shows use with transformers; llama.cpp's multimodal docs do not list this model. The Ollama Q4_K_M file for the instruct build is about 54.6 GB, so a 48 GB RTX A6000 cannot hold it. Recommended: A100 80GB for full Q4_K_M + vision. This is a base model — use for fine-tuning, few-shot completion, or as a vision backbone. Not suitable for direct chat without instruction tuning. For chat, use the instruct variant. vLLM dropped this architecture after v0.10.2. Few-shot prompting with image+text pairs is the primary base-model interaction pattern.
Hardware guidance
Minimum: about 57 GB of memory for Q4_K_M at 8K context, more than a 48 GB RTX A6000 or dual RTX 3090 holds. Recommended: A100 80GB at AWQ-INT4 for vision serving. VRAM math: identical to instruct variant. Q4 file ≈ 54.6 GB, KV cache at 8K ≈ 2.7 GB at FP16. Total at 8K: ~57 GB. A100 80GB: comfortable. Mac Studio M3 Ultra 96GB or 256GB: Q4_K_M + vision. RTX 5090 32GB: Q4_K_M only with partial CPU offload. Ollama lists only the instruct builds.
What breaks first
- Base model, not chat. Generates completions in whatever style the prompt sets. Feeding a conversational prompt produces unpredictable continuations unless formatted as few-shot with completion markers. 2. Few-shot context contamination. Base models are sensitive to prompt formatting. Extra whitespace or inconsistent few-shot formatting can degrade output quality dramatically. 3. Fine-tuning data format. Fine-tuning this model requires multimodal dataset formatting (text + image interleaving). Standard text-only fine-tuning scripts may not handle image tokens correctly.
Runtime recommendation
Hugging Face transformers, as shown on Meta's model card, for local base model vision use. vLLM dropped this architecture after v0.10.2, and llama.cpp's multimodal docs do not list it. For fine-tuning: Axolotl with multimodal support, or Meta's reference training code. Avoid Ollama for base vision models — Ollama is designed for instruct/chat models.
Common beginner mistakes
Mistake: Chatting with the base model and wondering why responses are nonsensical. Fix: Base models complete text, they don't follow instructions. Use few-shot completion format or switch to the instruct variant. Mistake: Assuming base and instruct have identical image understanding quality. Fix: Instruction tuning affects vision understanding. The base model's raw vision representations may differ from instruct. Test your task on both. Mistake: Fine-tuning without multimodal data format. Fix: The vision encoder expects image tokens in a specific format. Use a multimodal fine-tuning framework, not standard text-only scripts.
Family & lineage
How this model relates to others in its lineage. Family members share architecture and training-data roots; parent / children edges record direct distillation or fine-tune relationships.
Strengths
- Frontier-tier multimodal
- Strong visual reasoning
Weaknesses
- 64GB+ VRAM tier
Quantization variants
Each quantization trades model quality for file size and VRAM. Q4_K_M is the most popular starting point.
| Quantization | File size | VRAM required |
|---|---|---|
| AWQ-INT4 | 45.0 GB | 64 GB |
At Q4 Llama 3.2 90B Vision needs about 82 GB of memory by our estimate, more than a 32 GB graphics card holds. Renting a GPU with enough memory (or several GPUs) by the hour is the practical way to try it before you plan hardware for it.
Referral links: RunLocalAI may earn a commission or account credits if you sign up and rent, at no extra cost to you. Hourly rates and which GPUs are listed change all the time, so check the provider's page before you start. How we make money.
Get the model
HuggingFace
Original weights
Source repository — direct quantization required.
Hardware that runs this
Cards with enough VRAM for at least one quantization of Llama 3.2 90B Vision.
Models worth comparing
Same parameter band, plus what's one tier above and below — so you can decide what actually fits your hardware.
Frequently asked
What's the minimum VRAM to run Llama 3.2 90B Vision?
Can I use Llama 3.2 90B Vision commercially?
What's the context length of Llama 3.2 90B Vision?
Does Llama 3.2 90B Vision support images?
Source: huggingface.co/meta-llama/Llama-3.2-90B-Vision
Reviewed by RunLocalAI Editorial. See our editorial policy for how we research and verify model claims.
Related — keep moving
Verify Llama 3.2 90B Vision runs on your specific hardware before committing money.