Llama 3.2 90B Vision Instruct — Hardware Requirements & GPU Compatibility
VisionFunctions
Llama 3.2 90B Vision Instruct is a vision-language model in Meta's Llama 3 family, with around 89 billion parameters, that accepts image input alongside text and is tuned to follow chat-style instructions. It extends the Llama 3 lineup with multimodal understanding, aimed at tasks that combine visual and textual reasoning, such as describing images or answering questions about their content. It was released in September 2024 under Meta's Llama 3.2 community license, which permits use under Meta's specific terms rather than a fully open license. At around 89 billion parameters, local deployment at 4-bit precision needs a multi-GPU workstation or a machine with generous unified memory; most people will run it through a hosted API instead.
Llama 3.2 90B Vision Instruct runs on a 64 GB+ Mac or a workstation GPU
Best local pick → Mac Pro M2 Ultra (64 GB) · Q4_K_M · 63.5 GB · ~10 tok/s
Macs from 64 GB (Q4_K_M · ~10 tok/s)
llama3.2
Not set up for it?
Reference GPU · RTX 4090 24 GB
Estimates, ±20% · how we estimate
Run Llama 3.2 90B Vision Instruct
One recommended way. Every other method, OS and quant is in the step-by-step guide.Each way opens a step-by-step guide for your OS and quant.
Recommended
ollama run llama3.2-vision:90bQ4_K_M · needs ~63.5 GB · 16K context
Download: Weights on Hugging Face
Ways to runOther ways
each opens step-by-stepRun in cloud
Runs on A100 80GB (16 GB headroom) · Q4_K_M
- Generation speed
- ~23 tok/s
- generation speed
- Cost per 1M output tokens
- $13.90
- per 1M output tokens
How Much VRAM Does Llama 3.2 90B Vision Instruct Need?
Select a quantization level; the listed hardware adheres to it.
| Quantization | VRAM | +32K | On RTX 4090 (reference) | |
|---|---|---|---|---|
| 41.4 GB | 52.2 GB | 2-bit · K-quant — 2-bit quantization with K-quant improvements | Does not fit: needs 46.5 GB | |
| 47.5 GB | 58.3 GB | 3-bit medium — 3-bit medium quantization | Does not fit: needs 52.5 GB | |
| 58.5 GB | 69.3 GB | 4-bit medium — 4-bit medium quantization — most popular sweet spot | Does not fit: needs 63.5 GB | |
| 69.4 GB | 80.2 GB | 5-bit medium — 5-bit medium quantization — good quality/size tradeoff | Does not fit: needs 74.5 GB | |
| 80.4 GB | 91.2 GB | 6-bit, very good — 6-bit quantization, very good quality | Does not fit: needs 85.4 GB | |
| 97.5 GB | 108.2 GB | 8-bit, near-lossless — 8-bit quantization, near-lossless | Does not fit: needs 102.5 GB | |
| 194.9 GB | 205.7 GB | 16-bit · bfloat — Brain floating point 16 — preferred for training | Does not fit: needs 199.9 GB |
Selected Q4_K_M
4-bit medium quantization63.5GBat 16K context
- Weights
- 57.5 GB
- KV cache
- 5.8 GB
- Overhead
- 0.3 GB
Doesn't fit RTX 4090 (reference) · needs 39.5 GB more
Or .
Which Hardware Can Run Llama 3.2 90B Vision Instruct?
Q4_K_M · 58.5 GBLlama 3.2 90B Vision Instruct (Q4_K_M) requires 58.5 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 77+ GB is recommended. No consumer GPU has enough memory. 52 devices with unified memory can run Llama 3.2 90B Vision Instruct, including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB), Mac Pro M2 Ultra (64 GB).
48 GB
48 GB tier →Q4_K_M · best fit Q2_K (46.5 GB)
2 Mac classes · 1 device family
M5 Max48 GB~10Simulate output on Mac Studio M5 Max (48 GB)M5 Pro48 GB~5Simulate output on Mac Mini M5 Pro (48 GB)Snapdragon X2 Elite Extreme (X2E-96-100)~4Simulate output on Snapdragon X2 Elite Extreme Copilot+ PC64–80 GB
fits · 0.5 GB free · up to 16K ctx
3 Mac classes · 1 device family
M2 Ultra*64 GB~10Simulate output on Mac Pro M2 Ultra (64 GB)M5 Max64 GB~7Simulate output on Mac Studio M5 Max (64 GB)Orin~2Simulate output on NVIDIA Jetson AGX Orin 64GB
By VRAM tier · typical GPUs and Macs at Q4_K_M
How good is Llama 3.2 90B Vision Instruct?
llmrun Score blends public benchmark boards into one number; 100 = the average open model.
Not enough data yet for an llmrun Score.
A Score needs results on enough independent boards. How the Composite Score is built
Benchmarks
Here are all 5 boards that have results for this model, sorted by their weight in the llmrun Score; badges show who ran the test.
- GPQA Diamond
Epoch AI · Sep 2024Independent
Independent41.0%
#70/100 open
- AIME 2024/2025
Epoch AI · Sep 2024Independent
Independent2.6%
#75/88 open
- MATH Level 5
Epoch AI · Sep 2024Independent
Independent39.4%
#21/39 open
- MMLU
Epoch AI · Sep 2024Lab-reported
Lab-reported80.3%
#9/85 open
- BALROG
Epoch AI · Sep 2024Independent
Independent27.3%
#3/15 open
Alternatives to Llama 3.2 90B Vision Instruct
Llama 3 family · 65 models →Specifications
- Publisher
- Meta
- Family
- Llama 3
- Parameters
- 88.6B
- Release Date
- 2024-09-19
- License
- llama3.2
Frequently Asked Questions
How much VRAM does Llama 3.2 90B Vision Instruct need?
Llama 3.2 90B Vision Instruct requires 58.5 GB of VRAM at Q4_K_M, or 194.9 GB at BF16.
VRAM = Weights + KV Cache + Overhead
Weights = 88.6B × 4.8 bits ÷ 8 = 53.2 GB
KV Cache + Overhead ≈ 5.3 GB (at 2K context + ~0.3 GB framework)
Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.
VRAM usage by quantization
Can NVIDIA GeForce RTX 5090 run Llama 3.2 90B Vision Instruct?
No — Llama 3.2 90B Vision Instruct requires at least 41.4 GB at Q2_K, which exceeds the NVIDIA GeForce RTX 5090's 32 GB of VRAM.
What's the best quantization for Llama 3.2 90B Vision Instruct?
For Llama 3.2 90B Vision Instruct, Q4_K_M (58.5 GB) offers the best balance of quality and VRAM usage. Q5_K_M (69.4 GB) provides better quality if you have the VRAM. The smallest option is Q2_K at 41.4 GB.
VRAM requirement by quantization
★ Recommended — best balance of quality and VRAM usage.
Can I run Llama 3.2 90B Vision Instruct on a Mac?
Yes — Mac Pro M2 Ultra (64 GB) and 41 other Macs can run Llama 3.2 90B Vision Instruct. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 58.5 GB of usable unified memory (RAM minus macOS overhead).
Can I run Llama 3.2 90B Vision Instruct locally?
Yes — Llama 3.2 90B Vision Instruct can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.
How fast is Llama 3.2 90B Vision Instruct?
At Q4_K_M, Llama 3.2 90B Vision Instruct can reach ~82 tok/s on AMD Instinct MI350X. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.
tok/s = (bandwidth GB/s ÷ model GB) × efficiency
Example: NVIDIA B200 → 8000 ÷ 58.5 × 0.65 = ~89 tok/s
Estimated speed at Q4_K_M (58.5 GB)
Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.
What's the download size of Llama 3.2 90B Vision Instruct?
At Q4_K_M, the download is about 53.16 GB. The full-precision BF16 version is 177.19 GB. The smallest option (Q2_K) is 37.65 GB.
Which GPUs can run Llama 3.2 90B Vision Instruct?
No single consumer GPU has enough VRAM to run Llama 3.2 90B Vision Instruct at Q4_K_M (58.5 GB). Multi-GPU or professional hardware is required.
Which devices can run Llama 3.2 90B Vision Instruct?
52 devices with unified memory can run Llama 3.2 90B Vision Instruct at Q4_K_M (58.5 GB), including ASUS Ascent GX10, Beelink GTR9 Pro (Ryzen AI Max+ 395, 128 GB), Framework Desktop (Ryzen AI Max+ 395, 128 GB), GMKtec EVO-X2 (Ryzen AI Max+ 395, 128 GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.